Accepted Papers

---

Reinforcement learning usually makes use of numerical rewards, which have nice properties but also come with drawbacks and difficulties.Using rewards on an ordinal scale (ordinal rewards) is an alternative to numerical rewards that has received more attention in recent years.In this paper, a general approach to adapting reinforcement learning problems to the use of ordinal rewards is presented and motivated.We show how to convert common reinforcement learning algorithms to an ordinal variation by the example of Q-learning and introduce Ordinal Deep Q-Networks, which adapt deep reinforcement learning to ordinal rewards.Additionally, we run evaluations on problems provided by the OpenAI Gym framework, showing that our ordinal variants exhibit a performance that is comparable tothe numerical variations for a number of problems.We also give first evidence that our ordinal variant is able to produce better results for problems with less engineered and simpler-to-design reward signals.

---

---

While deep generative networks can simulate from complex data distributions, their utility can be hindered by limitations on the data available for training. Specifically, the training data distribution may differ from the target sampling distribution due to sample selection bias, or because the training data comes from a different but related distribution. We present methods to accommodate this difference via importance weighting, which allow us to estimate a loss function with respect to a target distribution even if we cannot access that distribution directly. These estimators, which differentially weight the contribution of data to the loss function, offer theoretical guarantees that heuristic approaches lack, while giving impressive empirical performance in a variety of settings.

---

---

Heterogeneous information networks (HINs) with rich semantics are ubiquitous in real-world applications. For a given HIN, many reasonable clustering results with distinct semantic meaning can simultaneously exist.User-guided clustering is hence of great practical value for HINs where users provide labels to a small portion of nodes.To cater to a broad spectrum of user guidance evidenced by different expected clustering results, carefully exploiting the signals residing in the data is potentially useful.Meanwhile, as one type of complex networks, HINs often encapsulate higher-order interactions that reflect the interlocked nature among nodes and edges.Network motifs, sometimes referred to as meta-graphs, have been used as tools to capture such higher-order interactions and reveal the many different semantics.We therefore approach the problem of user-guided clustering in HINs with network motifs.In this process, we identify the utility and importance of directly modeling higher-order interactions without collapsing them to pairwise interactions.To achieve this, we comprehensively transcribe the higher-order interaction signals to a series of tensors via motifs and propose the MoCHIN model based on joint non-negative tensor factorization.This approach applies to arbitrarily many, arbitrary forms of HIN motifs. An inference algorithm with speed-up methods is also proposed to tackle the challenge that tensor size grows exponentially as the number of nodes in a motif increases.We validate the effectiveness of the proposed method on two real-world datasets and three tasks, and MoCHIN outperforms all baselines in three evaluation tasks under three different metrics.Additional experiments demonstrated the utility of motifs and the benefit of directly modeling higher-order information especially when user guidance is limited.

---

---

We propose a novel ranking model that combines the Bradley-Terry-Luce probability model with a nonnegative matrix factorization framework to model and uncover the presence of latent variables that influence the performance of top tennis players. We derive an efficient, provably convergent, and numerically stable majorization-minimization-based algorithm to maximize the likelihood of datasets under the proposed statistical model. The model is tested on datasets involving the outcomes of matches between 20 top male and female tennis players over 14 major tournaments for men (including the Grand Slams and the ATP Masters 1000) and 16 major tournaments for women over the past 10 years. Our model automatically infers that the surface of the court (e.g., clay or hard court) is a key determinant of the performances of male players, but less so for females. Top players on various surfaces over this longitudinal period are also identified in an objective manner.

---

---

Value-based reinforcement-learning algorithms provide state-of-the-art results in model-free discrete-action settings, and tend to outperform actor-critic algorithms. We argue that actor-critic algorithms are limited by their need for an on-policy critic. We propose Bootstrapped Dual Policy Iteration (BDPI), a novel model-free reinforcement-learning algorithm for continuous states and discrete actions, with an actor and several off-policy critics. Off-policy critics are compatible with experience replay, ensuring high sample-efficiency, without the need for off-policy corrections. The actor, by slowly imitating the average greedy policy of the critics, leads to high-quality and state-specific exploration, which we compare to Thompson sampling. Because the actor and critics are fully decoupled, BDPI is remarkably stable, and unusually robust to its hyper-parameters. BDPI is significantly more sample-efficient than Bootstrapped DQN, PPO, and ACKTR, on discrete, continuous and pixel-based tasks.

---

---

We present the optimal weighted average (OWA) distributed learning algorithm for linear models. OWA achieves statistically optimal learning rates,uses only one round of communication,works on non-convex problems,and supports a fast cross validation procedure.The OWA algorithm first trains local models on each of the compute nodes;then a master machine merges the models using a second round of optimization. This second optimization uses only a small fraction of the data, and so has negligible computational cost.Compared with similar distributed estimators that merge locally trained models,OWA either has stronger statistical guarantees, is applicable to more models,or has a more computationally efficient merging procedure.

---

---

We strive to find contexts (i.e., subgroups of entities) under which exceptional (dis-)agreement occurs among a group of individuals, in any type of data featuring individuals (e.g., parliamentarians, customers) performing observable actions (e.g., votes, ratings) on entities (e.g., legislative procedures, movies). To this end, we introduce the problem of discovering statistically significant exceptional contextual intra-group agreement patterns. To handle the sparsity inherent to voting and rating data, we use Krippendorff's Alpha measure for assessing the agreement among individuals. We devise a branch-and-bound algorithm, named DEvIANT, to discover such patterns. DEvIANT exploits both closure operators and tight optimistic estimates. We derive analytic approximations for the confidence intervals (CIs) associated with patterns for a computationally efficient significance assessment. We prove that these approximate CIs are nested along specialization of patterns. This allows to incorporate pruning properties in DEvIANT to quickly discard non-significant patterns. Empirical study on several datasets demonstrates the efficiency and the usefulness of DEvIANT.

---

---

The area of constrained clustering has been extensively explored by researchers and used by practitioners. Constrained clustering formulations exist for popular algorithms such as k-means, mixture models, and spectral clustering but have several limitations. A fundamental strength of deep learning is its flexibility, and here we explore a deep learning framework for constrained clustering and in particular explore how it can extend the field of constrained clustering. We show that our framework can not only handle standard together/apart constraints (without the well documented negative effects reported earlier) generated from labeled side information but more complex constraints generated from new types of side information such as continuous values and high-level domain knowledge.

---

---

Existing methods for geolocating images use standard classification or image retrieval techniques. These methods have poor theoreticalproperties because they do not take advantage of the earth'sspherical geometry. In some cases, they require training data sets thatgrow exponentially with the number of feature dimensions. This paperintroduces the first image geolocation method that exploits the earth'sspherical geometry. Our method is based on the Mixture of von-MisesFisher (MvMF) distribution, which is a spherical analogue of the popularGaussian mixture model. We prove that this method requires only adataset of size linear in the number of feature dimensions, and empiricalresults show that our method outperforms previous methods with ordersof magnitude less training data and computation.

---

---

We present a new unsupervised method for learning general-purpose sentence embeddings.Unlike existing methods which rely on local contexts,such as words inside the sentence or immediately neighboring sentences,our method selects, for each target sentence,influential sentences from the entire document based on the document structure.We identify a dependency structure of sentences using metadata and text styles.Additionally, we propose an out-of-vocabulary word handling techniquefor the neural network outputs to model many domain-specific termswhich were mostly discarded by existing sentence embedding training methods.We empirically show that the model relies on the proposed dependenciesmore than the sequential dependency in many cases.We also validate our model on several NLP tasksshowing 23

---

---

String data are often disseminated to support applications such as location-based service provision or DNA sequence analysis. This dissemination, however, may expose sensitive patterns that model confidential knowledge (e.g., trips to mental health clinics from a string representing a user's location history). In this paper, we consider the problem of sanitizing a string by concealing the occurrences of sensitive patterns, while maintaining data utility. First, we propose a time-optimal algorithm, TFS-ALGO, to construct the shortest string preserving the order of appearance and the frequency of all non-sensitive patterns. Such a string allows accurately performing tasks based on the sequential nature and pattern frequencies of the string. Second, we propose a time-optimal algorithm, PFS-ALGO, which preserves a partial order of appearance of non-sensitive patterns but produces a much shorter string that can be analyzed more efficiently. The strings produced by either of these algorithms may reveal the location of sensitive patterns. In response, we propose a heuristic, MCSR-ALGO, which replaces letters in these strings with carefully selected letters, so that sensitive patterns are not reinstated and occurrences of spurious patterns are prevented. We implemented our sanitization approach that applies TFS-ALGO, PFS-ALGO and then MCSR-ALGO and experimentally show that it is effective and efficient.

---

---

Kernel two-sample testing is a useful statistical tool in determining whether data samples arise from different distributions without imposing any parametric assumptions on those distributions. However, raw data samples can expose sensitive information about individuals who participate in scientific studies, which makes the current tests vulnerable to privacy breaches. Hence, we design a new framework for kernel two-sample testing conforming to differential privacy constraints, in order to guarantee the privacy of subjects in the data. Unlike existing differentially private parametric tests that simply add noise to data, kernel-based testing imposes a challenge due to a complex dependence of test statistics on the raw data, as these statistics correspond to estimators of distances between representations of probability measures in Hilbert spaces. Our approach considers finite dimensional approximations to those representations. As a result, a simple chi-squared test is obtained, where a test statistic depends on a mean and covariance of empirical differences between the samples, which we perturb for a privacy guarantee. We investigate the utility of our framework in two realistic settingsand conclude that our method requires only a relatively modest increase in sample size to achieve a similar level of power to the non-private tests in both settings.

---

---

In this paper, we first identify activation shift, a simple but remarkablephenomenon in aneural network in whichthe preactivation value of a neuron has non-zero mean that depends on the anglebetween the weight vector of the neuron and the mean of the activationvector in the previous layer. We then propose linearly constrained weights (LCW) to reducethe activation shift in both fully connected and convolutional layers.The impact of reducing the activation shift in a neural network is studiedfrom the perspective of how the variance of variables in the networkchanges through layer operations in both forward and backward chains.We also discuss its relationship to the vanishing gradient problem.Experimental results show that LCW enables a deep feedforward networkwith sigmoid activation functions to be trained efficientlyby resolving the vanishing gradient problem.Moreover, combined with batch normalization, LCW improvesgeneralization performance of both feedforward andconvolutional networks.

---

---

The redundant sources problem in multi-source learning always exists in various real-world applications such as multimedia analysis, information retrieval, and medical diagnosis, in which the heterogeneous representations from different sources always have three-way redundancies. More seriously, the redundancies will cost a lot of storage space, cause high computational time, and degrade the performance of learner. This paper is an attempt to jointly reduce redundant sources. Specifically, a novel Heterogeneous Manifold Smoothness Learning (HMSL) model is proposed to linearly map multi-source data to a low-dimensional feature-isomorphic space, in which the information-correlated representations are close along manifold while the semantic-complementary instances are close in Euclidean distance. Furthermore, to eliminate three-way redundancies, we present a new Correlation-based Multi-source Redundancy Reduction (CMRR) method with 2,1-norm equation and generalized elementary transformation constraints to reduce redundant sources in the learned feature-isomorphic space. Comprehensive empirical investigations are presented that confirm the promise of our proposed framework.

---

---

We introduce a new Large Margin Gaussian Process (LMGP) model by formulating a pseudo-likelihood for a generalised multi-class hinge loss.We derive a highly scalable training objective for the proposed model using variational-inference and inducing point approximation.Additionally, we consider the joint learning of LMGP-DNN which combines the proposed model with traditional Deep Learning methods to enable learning for unstructured data.We demonstrate the effectiveness of the Large Margin GP with respect to both training time and accuracy in an extensive classification experiment consisting of 68 structured and two unstructured data sets.Finally, we highlight the key capability and usefulness of our model in yielding prediction uncertainty for classification by demonstrating its effectiveness in the tasks of large-scale active learning and detection of adversarial images.

---

---

Confounding bias, missing data, and selection bias are three common obstacles to valid causal inference in the data sciences. Covariate adjustment is the most pervasive technique for recovering casual effects from confounding bias. In this paper we introduce a covariate adjustment formulation for controlling confounding bias in the presence of missing-not-at-random data and develop a necessary and sufficient condition for recovering causal effects using the adjustment. We also introduce an adjustment formulation for controlling both confounding and selection biases in the presence of missing data and develop a necessary and sufficient condition for valid adjustment. Furthermore, we present an algorithm that lists all valid adjustment sets and an algorithm that finds a valid adjustment set containing the minimum number of variables, which are useful for researchers interested in selecting adjustment sets with desired properties.

---

---

In the last few years, many different methods have been focusing on using deep recurrent neural networks for natural language generation. The most widely used sequence-to-sequence neural methods are word-based: as such, they need a pre-processing step called delexicalization (conversely, relexicalization) to deal with uncommon or unknown words. These forms of processing, however, give rise to models that depend on the vocabulary used and are not completely neural.In this work, we present an end-to-end sequence-to-sequence model with attention mechanism which reads and generates at a character level, no longer requiring delexicalization, tokenization, nor even lowercasing. Moreover, since characters constitute the common "building blocks" of every text, it also allows a more general approach to text generation, enabling the possibility to exploit transfer learning for training.These skills are obtained thanks to two major features: (*) the possibility to alternate between the standard generation mechanism and a copy one, which allows to directly copy input facts to produce outputs, and(*) the use of an original training pipeline that further improves the quality of the generated texts.We also introduce a new dataset called E2E+, designed to highlight the copying capabilities of character-based models, that is a modified version of the well-known E2E dataset used in the E2E Challenge. We tested our model according to five broadly accepted metrics (including the widely used bleu), showing that it yields competitive performance with respect to both character-based and word-based approaches.

---

---

Hierarchical clustering is a fundamental tool in data mining, machine learning and statistics. Popular hierarchical clustering algorithms include top-down divisive approaches such as bisecting k-means, k-median, and k-center and bottom-up agglomerative approaches such as single-linkage, average-linkage, and centroid-linkage. Unfortunately, only a few scalable hierarchical clustering algorithms are known, mostly based on the single-linkage algorithm. So, as datasets increase in size every day, there is a pressing need to scale other popular methods.We introduce efficient distributed algorithms for bisecting k-means, k-median, and k-center as well as centroid-linkage. In particular, we first formalize a notion of closeness for a hierarchical clustering algorithm, and then we use this notion to design new scalable distributed methods with strong worst case bounds on the running time and the quality of the solutions. Finally, we show experimentally that the introduced algorithms are efficient and close to their sequential variants in practice.

---

---

Given a labeled dataset that contains a rare (or minority) class of of-interest instances, as well as a large class of instances that are not of interest,how can we learn to recognize future of-interest instances over a continuous stream?We introduce RaRecognize, which (i) estimates a general decision boundary between the rare and the majority class, (ii) learns to recognize individual rare subclasses that exist within the training data, as well as (iii) flags instances from previously unseen rare subclasses as newly emerging.The learner in (i) is general in the sense that by construction it is dissimilar to the specialized learners in (ii), thus distinguishes minority from the majority without overly tuning to what is seen in the training data. Thanks to this generality, RaRecognize ignores all future instances that it labels as majorityand recognizes the recurrent as well as emerging rare subclasses only. This saves effort at test time as well as ensures that the model size grows moderately over time as it only maintains specialized minority learners.Through extensive experiments, we show that RaRecognize outperforms state-of-the art baselines on three real-world datasets that contain corporate-risk and disaster documents as rare classes.

---

---

Unsupervised anomaly detection is commonly performed using a distance or density based technique, such as K-Nearest neighbours, Local Outlier Factor or One-class Support Vector Machines. One-class Support Vector Machines reduce the computational cost of testing new data by providing sparse solutions. However, all these techniques have relatively high computational requirements for training. Moreover, identifying anomalies based solely on density or distance is not sufficient when both point (isolated) and cluster anomalies exist in an unlabelled training set. Finally, these unsupervised anomaly detection techniques are not readily adapted for active learning, where the training algorithm should identify examples for which labelling would make a significant impact on the accuracy of the learned model. In this paper, we propose a novel technique called Maximin-based Anomaly Detection that addresses these challenges by selecting a representative subset of data in combination with a kernel-based model construction. We show that the proposed technique (a) provides a statistically significant improvement in the accuracy as well as the computation time required for training and testing compared to several benchmark unsupervised anomaly detection techniques, and (b) effectively uses active learning with a limited budget.

---

---

Deep learning is very effective at jointly learning feature representations and classification models, especially when dealing with high dimensional input patterns. Probabilistic logic reasoning, on the other hand, is capable of take consistent and robust decisions in complex environments. The integration of deep learning and logic reasoning is still an open-research problem and it is considered to be the key for the development of real intelligent agents. This paper presents Deep Logic Models, which are deep graphical models integrating deep learning and logic reasoning both for learning and inference. Deep Logic Models create an end-to-end differentiable architecture, where deep learners are embedded into a network implementing a continuous relaxation of the logic knowledge. The learning process allows to jointly learn the weights of the deep learners and the meta-parameters controlling the high-level reasoning. The experimental results show that the proposed methodology overcomes the limitations of the other approaches that have been proposed to bridge deep learning and reasoning.

---

---

In spite of the amazing results obtained by deep learning in many applications, a real intelligent behavior of an agent acting in a complex environment is likely to require some kind of higher-level symbolic inference. Therefore, there is a clear need for the definition of a general and tight integration between low-level tasks, processing sensorial data that can be effectively elaborated using deep learning techniques, and the logic reasoning that allows humans to take decisions in complex environments.This paper presents LYRICS, a generic interface layer for AI, which is implemented in TersorFlow (TF). LYRICS provides an input language that allows to define arbitrary First Order Logic (FOL) background knowledge. The predicates and functions of the FOL knowledge can be bound to any TF computational graph, and the formulas are converted into a set of real-valued constraints, which participate to the overall optimization problem. This allows to learn the weights of the learners, under the constraints imposed by the prior knowledge.The framework is extremely general as it imposes no restrictions in terms of which models or knowledge can be integrated. In this paper, we show the generality of the approach showing some use cases of the presented language, including model checking, supervised learning and collective classification.

---

---

Maintaining an accurate trained model on an infinite data stream is challenging due to concept drifts that render a learned model inaccurate.Updating the model periodically can be expensive, and so traditional approaches for computationally limited devices involve a variation of online or incremental learning, which tend to be less robust.The advent of heterogeneous architectures and Internet-connected devices gives rise to a new opportunity.A weak processor can call upon a stronger processor or a cloud server to perform a complete batch training pass once a concept drift is detected - trading power or network bandwidth for increased accuracy.We capitalize on this opportunity in two steps.We first develop a computationally efficient bound for changes in any linear model with convex, differentiable loss.We then propose a sliding window-based algorithm that uses a small number of batch model computations to maintain an accurate model of the data stream. It uses the bound to continuously evaluate the difference between the parameters of the existing model and a hypothetical optimal model, triggering computation only as needed.Empirical evaluation on real and synthetic datasets shows that our proposed algorithm adapts well to concept drifts and provides a better tradeoff between the number of model computations and model accuracy than classic concept drift detectors.When predicting changes in electricity prices, for example, we achieve 6

---

---

In the densest subgraph problem, given an undirected graph G(V,E,w) with non-negative edge weights we are asked to find a set of nodes S⊆ V that maximizes the degree density w(S)/|S|, where w(S) is the sum of the weights of the edges in the graph induced by S. This problem is solvable in polynomial time, and in general is well studied. But what happens when the edge weights are negative? Is the problem still solvable in polynomial time? Also, why should we care about the densest subgraph problem in the presence of negative weights? In this work we answer the aforementioned questions. Specifically, we provide two novel graph mining primitives that are applicable to a wide variety of applications. Our primitives can be used to answer questions such as "how can we find a dense subgraph in Twitter with lots of replies and mentions but no follows?", "how do we extract a dense subgraph with high expected reward and low risk from an uncertain graph"? We formulate both problems mathematically as special instances of dense subgraph discovery in graphs with negative weights. We study the hardness of the problem, and we prove that the problem in general is NP-hard, but we also provide sufficient conditions under which it is poly-time solvable. We design an efficient approximation algorithm that works best in the presence of small negative weights, and an effective heuristic for the more general case. Finally, we perform experiments on various real-world datasets that verify the value of the proposed primitives, and the effectiveness of our proposed algorithms.

---

---

Dimensionality reduction (DR) on the manifold includes effectivemethods which project the data from an implicit relational space ontoa vectorial space. Regardless of the achievements in this area, these algorithmssuffer from the lack of interpretation of the projection dimensions.Therefore, it is often difficult to explain the physical meaning behindthe embedding dimensions. In this research, we propose the interpretablekernel DR algorithm (I-KDR) as a new algorithm which maps the datafrom the feature space to a lower dimensional space where the classes aremore condensed with less overlapping. Besides, the algorithm creates thedimensions upon local contributions of the data samples, which makes iteasier to interpret them by class labels. Additionally, we efficiently fusethe DR with feature selection task to select the most relevant features ofthe original space to the discriminative objective. Based on the empiricalevidence, I-KDR provides better interpretations for embedding dimensionsas well as higher discriminative performance in the embedded spacecompared to the state-of-the-art and popular DR algorithms.

---

---

Feature selection is central to modern data science. The `stability' of a feature selection algorithm refers to the sensitivity of its choices to small changes in training data. This is, in effect, the robustness of the chosen features. This paper considers the estimation of stability when we expect strong pairwise correlations, otherwise known as feature redundancy. We demonstrate that existing measures are inappropriate here, as they systematically underestimate the true stability, giving an overly pessimistic view of a feature set.We propose a new statistical measure which overcomes this issue, and generalises previous work.

---

---

This paper introduces the Elliptical Basis Function Data Descriptor (EBFDD) network, a one-class classification approach to anomaly detection based on Radial Basis Function (RBF) neural networks. The EBFDD network uses elliptical basis functions, which allows it to learn sophisticated decision boundaries while retaining the advantages of a shallow network. We have proposed a novel cost function, whose minimisation results in a trained anomaly detector that only requires examples of the normal class at training time. The paper includes a large benchmark experiment that evaluates the performance of EBFDD network and compares it to state of the art one-class classification algorithms including the One-Class Support Vector Machine and the Isolation Forest. The experiments show that, overall, the EBFDD network outperforms the state of the art approaches.

---

---

Using touch devices to navigate in virtual 3D environments such as computer assisted design (CAD) models or geographical information systems (GIS) is inherently difficult for humans, as the 3D operations have to be performed by the user on a 2D touch surface. This ill-posed problem is classically solved with a fixed and handcrafted interaction protocol, which must be learned by the user. We propose to automatically learn a new interaction protocol allowing to map a 2D user input to 3D actions in virtual environments using reinforcement learning (RL). A fundamental problem of RL methods is the vast amount of interactions often required, which are difficult to come by when humans are involved. To overcome this limitation, we make use of two collaborative agents. The first agent models the human by learning to perform the 2D finger trajectories. The second agent acts as the interaction protocol, interpreting and translating to 3D operations the 2D finger trajectories from the first agent. We restrict the learned 2D trajectories to be similar to a training set of collected human gestures by first performing state representation learning, prior to reinforcement learning. This state representation learning is addressed by projecting the gestures into a latent space learned by a variational auto encoder (VAE).

---

---

Post-hoc interpretability approaches, although powerful tools to generate explanations for predictions made by a trained black-box model, have been shown to be vulnerable to issues caused by lack of robustness of the classifier. In particular, this paper focuses on the notion of explanation justification, defined as connectedness to ground-truth data, in the context of counterfactuals. In this work, we explore the extent of the risk of generating unjustified explanations. We propose an empirical study to assess the vulnerability of classifiers and show that the chosen learning algorithm heavily impacts the vulnerability of the model. Additionally, we show that state-of-the-art post-hoc counterfactual approaches can minimize the impact of this risk by generating less local explanations.

---

---

We study involuntary micro-movements of the eye for biometric identification. While prior studies extract lower-frequency macro-movements from the output of video-based eye-tracking systems and engineer explicit features of these macro-movements, we develop a deep convolutional architecture that processes the raw eye-tracking signal. Compared to prior work, the network attains a lower error rate by one order of magnitude and is faster by two orders of magnitude: it identifies users accurately within seconds.

---

---

We consider distributed on-device learning with limited communication and security requirements. We propose a new robust distributed optimization algorithm with efficient communication and attack tolerance. The proposed algorithm has provable convergence and robustness under non-IID settings. Empirical results show that the proposed algorithm stabilizes the convergence and tolerates data poisoning on a small number of workers.

---

---

In statistics and machine learning, approximation of an intractable integration is often achieved by using the unbiased Monte Carlo estimator, but the variances of the estimation are generally high in many applications. Control variates approaches are well-known to reduce the variance of the estimation. These control variates are typically constructed by employing predefined parametric functions or polynomials, determined by using those samples drawn from the relevant distributions. Instead, we propose to construct those control variates by learning neural networks to handle the cases when test functions are complex. In many applications, obtaining a large number of samples for Monte Carlo estimation is expensive, the adoption of the original loss function may result in severe overfitting when training a neural network. This issue was not reported in those literature on control variates with neural networks. We thus further introduce a constrained control variates with neural networks to alleviate the overfitting issue. We apply the proposed control variates to both toy and real data problems, including a synthetic data problem, Bayesian model evidence evaluation and Bayesian neural networks. Experimental results demonstrate that our method can achieve significant variance reduction compared to other methods.

---

---

We propose a novel approach for learning node representations in directed graphs, which maintains separate views or embedding spaces for the two distinct node roles induced by the directionality of the edges.We argue that the previous approaches either fail to encode the edge directionality or their encodings cannot be generalized across tasks.With our simple alternating random walk strategy, we generate role specific vertex neighborhoods and train node embeddings in their corresponding source/target roles while fully exploiting the semantics of directed graphs. We also unearth the limitations of evaluations on directed graphs in previous works and propose a clear strategy for evaluating link prediction and graph reconstruction in directed graphs. We conduct extensive experiments to showcase our effectiveness on several real-world datasets on link prediction, node classification and graph reconstruction tasks. We show that the embeddings from our approach are indeed robust, generalizable and well performing across multiple kinds of tasks and graphs. We show that we consistently outperform all baselines for node classification task. In addition to providing a theoretical interpretation of our method we also show that we are considerably more robust than the other directed graph approaches.

---

---

Learning domain-invariant representation is a dominant approach for domain generalization (DG), where we need to build a classifier that is robust toward domain shifts.However, previous domain-invariance-based methods overlooked the underlying dependency of classes on domains, which is responsible for the trade-off between classification accuracy and domain invariance.Because the primary purpose of DG is to classify unseen domains rather than the invariance itself, the improvement of the invariance can negatively affect DG performance under this trade-off.To overcome the problem, this study first expands the analysis of the trade-off by Xie et. al., and provides the notion of accuracy-constrained domain invariance, which means the maximum domain invariance within a range that does not interfere with accuracy.We then propose a novel method adversarial feature learning with accuracy constraint (AFLAC), which explicitly leads to that invariance on adversarial training.Empirical validations show that the performance of AFLAC is superior to that of domain-invariance-based methods on both synthetic and three real-world datasets, supporting the importance of considering the dependency and the efficacy of the proposed method.

---

---

The data association problem is concerned with separating data coming from different generating processes, for example when data comes from different data sources, contain significant noise, or exhibit multimodality.We present a fully Bayesian approach to this problem.Our model is capable of simultaneously solving the data association problem and the induced supervised learning problem.Underpinning our approach is the use of Gaussian process priors to encode the structure of both the data and the data associations.We present an efficient learning scheme based on doubly stochastic variational inference and discuss how it can be applied to deep Gaussian process priors.

---

---

Partial label learning (PLL) is a weakly supervised learning framework which learns from the data where each example is associated with a set of candidate labels, among which only one is correct. Most existing approaches are based on the disambiguation strategy, which either identifies the valid label iteratively or treats each candidate label equally based on the averaging strategy. In both cases, the disambiguation strategy shares a common shortcoming that the ground-truth label may be overwhelmed by the false positive candidate labels, especially when the number of candidate labels becomes large. In this paper, a probability propagation method for partial label learning (PP-PLL) is proposed. Specifically, based on the manifold assumption, a biconvex regular function is proposed to model the linear mapping relationships between input features and output true labels. In PP-PLL, the topological relations among training samples are used as additional information to strengthen the mutual exclusiveness among candidate labels, which helps to prevent the ground-truth label from being overwhelmed by a large number of candidate labels. Experimental studies on both artificial and real-world data sets demonstrate that the proposed PP-PLL method can achieve superior or comparable performance against the state-of-the-art methods.

---

---

Human eye gaze patterns are highly individually characteristic. Gaze patterns observed during the routine access of a user to a device or document can therefore be used to identify subjects unobtrusively,that is, without the need to perform an explicit verification such as entering a password.Existing approaches to biometric identification from gaze patterns segment raw gaze data into short, local patterns called saccades and fixations. Subjects are then identified by characterizing the distribution of these patterns or deriving hand-crafted features for them.In this paper, we follow a different approach by training deep neural networks directly on the raw gaze data.As the distribution of short, local patterns has been shown to be particularly informative for distinguishing subjects, we introduce a parameterized and end-to-end learnable statistical aggregation layer called the quantile layer that enables the network to explicitly fit the distribution offilter activations in preceeding layers. We empirically show that deep neural networks with quantile layers outperform existing probabilistic and feature-based methods for identifying subjectsbased on eye movements by a large margin.

---

---

T-distributed stochastic neighbour embedding (t-SNE) is a widely used data visualisation technique. It differs from its predecessor SNE by the low-dimensional similarity kernel: the Gaussian kernel was replaced by the heavy-tailed Cauchy kernel, solving the `crowding problem' of SNE. Here, we develop an efficient implementation of t-SNE for a t-distribution kernel with an arbitrary degree of freedom ν, with ν→∞ corresponding to SNE and ν=1 corresponding to the standard t-SNE. Using theoretical analysis and toy examples, we show that ν<1 can further reduce the crowding problem and reveal finer cluster structure that is invisible in standard t-SNE. We further demonstrate the striking effect of heavier-tailed kernels on large real-life data sets such as MNIST, single-cell RNA-sequencing data, and the HathiTrust library. We use domain knowledge to confirm that the revealed clusters are meaningful. Overall, we argue that modifying the tail heaviness of the t-SNE kernel can yield additional insight into the cluster structure of the data.

---

---

We present a multistage procedure to cluster directed and undirected weighted graphs by finding the block structure of their adjacency matrices. A central part of the process is to scale the adjacency matrix into a doubly-stochastic form, which permits detection of the whole matrix block structure with minimal spectral information (theoretically a single pair of singular vectors suffices).We present the different stages of our method, namely the impact of the doubly-stochastic scaling on singular vectors, detection of the block structure by means of these vectors, and details such as cluster refinement and a stopping criterion. Then we test thealgorithm's effectiveness by using it on two unsupervised classification tasks: community detection in networks and shape detection in cloudsof points in two dimensions. By comparing results of our approach with thoseof widely used algorithms designed for specific purposes, we observe that our method is competitive (for community detection) if not superior (for shape detection) in comparison with existing methods.

---

---

Batch Reinforcement Learning (Batch RL) consists in training a policy using trajectories collected with another policy, called the behavioural policy. Safe policy improvement (SPI) provides guarantees with high probability that the trained policy performs better than the behavioural policy, also called baseline in this setting. Previous work shows that the SPI objective improves mean performance as compared to using the basic RL objective, which boils down to solving the MDP with maximum likelihood. Here, we build on that work and improve more precisely the SPI with Baseline Bootstrapping algorithm (SPIBB) by allowing the policy search over a wider set of policies. Instead of binarily classifying the state-action pairs into two sets (the uncertain and the safe-to-train-on ones), we adopt a softer strategy that controls the error in the value estimates by constraining the policy change according to the local model uncertainty. The method can take more risks on uncertain actions all the while remaining provably-safe, and is therefore less conservative than the state-of-the-art methods. We propose two algorithms (one optimal and one approximate) to solve this constrained optimization problem and empirically show a significant improvement over existing SPI algorithms both on finite MDPS and on infinite MDPs with a neural network function approximation.

---

---

Much of human knowledge is encoded in text, available in scientific publications, books, and the web. Given the rapid growth of these resources, we need automated methods to extract such knowledge into machine-processable structures, such as knowledge graphs. An important task in this process is entity normalization, which consists of mapping noisy entity mentions in text to canonical entities in well-known reference sets. However, entity normalization is a challenging problem; there often are many textual forms for a canonical entity that may not be captured in the reference set, and entities mentioned in text may include many syntactic variations, or errors. The problem is particularly acute in scientific domains, such as biology. To address this problem, we have developed a general, scalable solution based on a deep Siamese neural network model to embed the semantic information about the entities, as well as their syntactic variations. We use these embeddings for fast mapping of new entities to large reference sets, and empirically show the effectiveness of our framework in challenging bio-entity normalization datasets.

---

---

Label ranking considers the problem of learning a mapping from instances to strict total orders over a predefined set of labels. In this paper, we present a framework for label ranking using a decomposition into a set of multiclass problems. Conceptually, our approach can be seen as a generalization of pairwise preference learning. In contrast to the latter, it allows for controlling the granularity of the decomposition, varying between binary preferences and complete rankings as extreme cases. It is specifically motivated by limitations of pairwise learning with regard to the minimization of certain loss functions. We discuss theoretical properties of the proposed method in terms of accuracy, error correction, and computational complexity. Experimental results are promising and indicate that improvements upon the special case of pairwise preference decomposition are indeed possible.

---

---

A multi-label classifier assigns a set of labels to each data object. A natural requirement in many end-use applications is that the classifier also provides a well-calibrated confidence (probability) to indicate the likelihood of the predicted set being correct; for example, an application may automate high-confidence predictions while manually verifying low-confidence predictions. The simplest multi-label classifier, called Binary Relevance (BR), applies one binary classifier to each label independently and takes the product of the individual label probabilities as the overall label-set probability (confidence). Despite its many known drawbacks, such as generating suboptimal predictions and poorly calibrated confidence scores, BR is widely used in practice due to its speed and simplicity. We seek in this work to improve both BR's confidence estimation and prediction through a post calibration and reranking procedure. We take the BR predicted set of labels and its product score as features, extract more features from the prediction itself to capture label constraints, and apply Gradient Boosted Trees (GB) as a calibrator to map these features into a calibrated confidence score. GB not only produces well-calibrated scores (aligned with accuracy and sharp), but also models label interactions, correcting a critical flaw in BR. We further show that reranking label sets by the new calibrated confidence makes accurate set predictions on par with state-of-the-art multi-label classifiers - yet calibrated, simpler, and faster.

---

---

We present a pairwise learning to rank approach based on a neural net, called DirectRanker, that generalizes the RankNet architecture. We show mathematically that our model is reflexive, antisymmetric, and transitive allowing for simplified training and improved performance. Experimental results on the LETOR MSLR-WEB10K, MQ2007 and MQ2008 datasets show that our model outperforms numerous state-of-the-art methods, while being inherently simpler in structure and using a pairwise approach only.

---

---

We consider the problem of online planning in a Markov Decision Process when given only access to a generative model, restricted to open-loop policies - i.e. sequences of actions - and under budget constraint. In this setting, the Open-Loop Optimistic Planning (OLOP) algorithm enjoys good theoretical guarantees but is overly conservative in practice, as we show in numerical experiments. We propose a modified version of the algorithm with tighter upper-confidence bounds, KL-OLOP, that leads to better practical performances while retaining the sample complexity bound. Finally, we propose an efficient implementation that significantly improves the time complexity of both algorithms.

---

---

Link prediction requires predicting which new links are likely to appear in a graph.In this paper, we present an approach for link prediction that relies on higher-order analysis of the graph topology, well beyond the typical approach which relies on common neighbors.We treat the link prediction problem as a supervised classification problem,and we propose a set of features that depend on the patterns or motifs that a pair of nodes occurs in.By using motifs of sizes 3, 4, and 5, our approach captures a high level of detail about the graph topology. In addition,we propose two optimizations to construct the classification dataset from the graph.First, we propose adding negative examples to the graph as an alternative to the common approach of removing positive ones.Second, we show that it is important to control for the shortest-path distance when sampling pairs of nodes to form negative examples, since the difficulty of prediction varies with the distance.We experimentally demonstrate that using our proposed motif features in off-the-shelf classifiers results in up to 10 percentage points increase in accuracy over prior topology-based and feature-learning methods.

---

---

Multitask algorithms typically use task similarity information as a bias to speed up and improve the performance of learning processes. Tasks are learned jointly, sharing information across them, inorder to construct models more accurate than those learned separatelyover single tasks. In this contribution, we present the first multitaskmodel, to our knowledge, based on Hopfield Networks (HNs), namedHoMTask. We show that by appropriately building a unique HN embedding all tasks, a more robust and effective classification model can be learned. HoMTask is a transductive semi-supervised parametric HN, thatminimizes an energy function extended to all nodes and to all tasks understudy. We provide theoretical evidence that the optimal parameters automatically estimated by HoMTask make coherent the model itself with the prior knowledge (connection weights and node labels). The convergence properties of HNs are preserved, and the fixed point reached by the network dynamics gives rise to the prediction of unlabeled nodes. The proposed model improves the classification abilities of singletask HNs on a preliminary benchmark comparison, and achievescompetitive performance with state-of-the-art semi-supervised graph-based algorithms.

---

---

We derive a tightened empirical Bernstein bound (EBB) on the variation of the sample mean from the population mean, and show that it improves the performance of upper confidence bound (UCB) methods in multi-armed bandit problems.Like other EBBs, our EBB is a concentration inequality for the variation of the sample mean in terms of the sample variance.Its derivation uses a combination of probability unions and Chernoff bounds for the mean of samples and mean of sample squares.Analysis reveals that our approach can tighten the best existing EBBs by about a third, and thereby halves the distance to a bound constructed with perfect variance information.We illustrate the practical usefulness of our novel EBB by applying it to a multi-armed bandit problem as a component of a UCB method. Our method outperforms existing approaches by producing lower expected regret than variants of UCB employing several other bounds, including state-of-the-art EBBs.

---

---

Multi-label classification (MLC) is the task of assigning a set of target labels for a given sample. Modeling the combinatorial label interactions in MLC has been a long-haul challenge. We propose Label Message Passing (LaMP) Neural Networks to efficiently model the joint prediction of multiple labels. LaMP treats labels as nodes on a label-interaction graph and computes the hidden representation of each label node conditioned on the input using attention-based neural message passing. Attention enables LaMP to assign different importances to neighbor nodes per label, learning how labels interact (implicitly). The proposed models are simple, accurate, interpretable, structure-agnostic, and applicable for predicting dense labels since LaMP is incredibly parallelizable. We validate the benefits of LaMP on seven real-world MLC datasets, covering a broad spectrum of input/output types and outperforming the state-of-the-art results. Notably, LaMP enables intuitive interpretation of how classifying each label depends on the elements of a sample and at the same time rely on its interaction with other labels.

---

---

Dense subtensor detection gains remarkable success in spotting anomaly and fraudulent behaviors for the multi-aspect data (i.e., tensors), like in social media and event streams.Existing methods detect the densest subtensors flatly and separately, with an underlying assumption that those subtensors are exclusive.However, many real-world tensors usually present hierarchical properties, e.g., the core-periphery structure or dynamic communities in networks. In this paper, we propose CatchCore, a novel framework to effectively find the hierarchical dense subtensors. We first design a unified metric for dense subtensor detection, which can be optimized with gradient-based methods. With the proposed metric, detects hierarchical dense subtensors through the hierarchy-wise alternative optimization.Finally, we utilize the minimum description length principle to measure the quality of detection result and select the optimal hierarchical dense subtensors.Extensive experiments on synthetic and real-world datasets demonstrate that outperforms the top competitors in accuracy for detecting dense subtensors and anomaly patterns. Additionally, CatchCore successfully identified a hierarchical researcher co-authorship group with intense interactions in DBLP dataset. Meanwhile, CatchCore also scales linearly with all aspects of tensors.

---

---

In this paper, we initiate the study of the problem of ordering objects from their pairwise comparison results when allowed to discard up to a certain number of objects as outliers. More specifically, we seek to find an ordering under the popular Kendall tau distance measure, i.e., minimizing the number of pairwise comparison results that are inconsistent with the ordering, with some outliers removed. The presence of outliers challenges the assumption that a global consistent ordering exists and obscures the measure. This problem does not admit a polynomial time algorithm unless NP ⊆ BPP, and therefore, we develop approximation algorithms with provable guarantees for all inputs. Our algorithms have running time and memory usage that are almost linear in the input size. Further, they are readily adaptable to run on massively parallel platforms such as MapReduce or Spark.

---

---

Social recommendation systems typically combine extra information like a social network with the user-item interaction network in order to alleviate data sparsity issues. This also helps in making more accurate and personalized recommendations. However, most of the existing systems work under the assumption that all socially connected users have equal influence on each other in a social network, which is not true in practice. Further, estimating the quantum of influence that exists among entities in a user-item interaction network is essential when only implicit ratings are available. This has been ignored even in many recent state-of-the-art models such as SAMN (Social Attentional Memory Network) and DeepSoR (Deep neural network model on Social Relations). Many a time, capturing a complex relationship between the entities (users/items) is essential to boost the performance of a recommendation system. We address these limitations by proposing a novel neural network model, SoRecGAT, which employs multi-head and multi-layer graph attention mechanism. The attention mechanism helps the model learn the influence of entities on each other more accurately. The proposed model also takes care of heterogeneity among the entities seamlessly. SoRecGAT is a general approach and we also validate its suitability when information in the form of a network of co-purchased items is available. Empirical results on eight real-world datasets demonstrate that the proposed model outperforms state-of-the-art models.

---

---

Minimizing expected loss measured by a proper scoring rule, such as Brier score or log-loss (cross-entropy), is a common objective while training a probabilistic classifier. If the data have experienced dataset shift where the class distributions change post-training, then often the model's performance will decrease, over-estimating the probabilities of some classes while under-estimating the others on average. We propose unbounded and bounded general adjustment (UGA and BGA) methods that transform all predictions to (re-)equalize the average prediction and the class distribution. These methods act differently depending on which proper scoring rule is to be minimized, and we have a theoretical guarantee of reducing loss on test data, if the exact class distribution is known. We also demonstrate experimentally that, when in practice the class distribution is known only approximately, there is often still a reduction in loss depending on the amount of shift and the precision to which the class distribution is known.

---

---

Stochastic elements in reinforcement learning (RL) have shown promise to improve exploration and handling of uncertainty, such as the utilization of stochastic weights in NoisyNets and stochastic policies in the maximum entropy RL frameworks. Yet effective and general approaches to include such elements in actor-critic models are still lacking. Inspired by the aforementioned techniques, we propose an effective way to inject randomness into actor-critic models to improve general exploratory behavior and reflect environment uncertainty. Specifically, randomness is added at the level of intermediate activations that feed into both policy and value functions to achieve better correlated and more complex perturbations. The proposed framework also features flexibility and simplicity, which allows straightforward adaptation to a variety of tasks. We test several actor-critic models enhanced with stochastic activations and demonstrate their effectiveness in a wide range of Atari 2600 games, a continuous control problem and a car racing task. Lastly, in a qualitative analysis, we present evidence of the proposed model adapting the noise in the policy and value functions to reflect uncertainty and ambiguity in the environment.

---

---

Given a decision forest, we study a problem of reducing the number of its distinct branching conditionswithout changing each tree's structure while keeping classification performance.A decision forest with a smaller number of distinct branching conditions can not only have a smaller description lengthbut also be implemented by hardware more efficiently. To force the modified decision forest to keep classification performance, we consider a conditionthat the decision paths at each branching node do not change for 100σ% of the given feature vectors passing through the node for a given 0≤σ<1.Under this condition, we propose an algorithm that minimizes the number of distinct branching conditions by sharing the same condition among multiple branching nodes.According to our experimental results using 13 datasets in UCI machine learning repository, our algorithm succeeded more than 90

---

---

Bag-of-Concepts, a model that counts the frequency of clustered word embeddings (i.e., concepts) in a document, has demonstrated the feasibility of leveraging clustered word embeddings to create features for document representation. However, information is lost as the word embeddings themselves are not used in the resulting feature vector. This paper presents a novel text representation method, Vectors of Locally Aggregated Concepts (VLAC). Like Bag-of-Concepts, it clusters word embeddings for its feature generation. However, instead of counting the frequency of clustered word embeddings, VLAC takes each cluster's sum of residuals with respect to its centroid and concatenates those to create a feature vector. The resulting feature vectors contain more discriminative information than Bag-of-Concepts due to the additional inclusion of these first order statistics. The proposed method is tested on four different data sets for single-label classification and compared with several baselines, including TF-IDF and Bag-of-Concepts. Results indicate that when combining features of VLAC with TF-IDF significant improvements in performance were found regardless of which word embeddings were used.

---

---

Graph signal processing is a useful tool for representing, analyzing, and processing the signal lying on a graph, and has attracted attention in several fields including data mining and machine learning.A key to construct the graph signal processing is the graph Fourier transform, which is defined by using eigenvectors of the graph Laplacian of an undirected graph.The orthonormality of eigenvectors gives the graph Fourier transform algebraically desirable properties, and thus the graph signal processing for undirected graphs has been well developed.However, since eigenvectors of the graph Laplacian of a directed graph are generally not orthonormal, it is difficult to simply extend the graph signal processing to directed graphs.In this paper, we present a general framework for extending the graph signal processing to directed graphs.To this end, we introduce the Hermitian Laplacian which is a complex matrix obtained from an extension of the graph Laplacian.The Hermitian Laplacian is defined so as to preserve the edge directionality and Hermitian property and enables the graph signal processing to be straightforwardly extended to directed graphs.Furthermore, the Hermitian Laplacian guarantees some desirable properties, such as non-negative real eigenvalues and the unitarity of the Fourier transform.Finally, experimental results for representation learning and signal denoising of/on directed graphs show the effectiveness of our framework.

---

---

Gaussian processes (GPs) are an elegant Bayesian approach to model an unknown function. The choice of the kernel characterizes one's assumption on how the unknown function autocovaries. It is a core aspect of a GP design, since the posterior distribution can significantly vary for different kernels. The spectral mixture (SM) kernel is derived by modelling a spectral density - the Fourier transform of a kernel - with a linear mixture of Gaussian components. As such, the SM kernel cannot model dependencies between components. In this paper we use cross convolution to model dependencies between components and derive a new kernel called Generalized Convolution Spectral Mixture (GCSM). Experimental analysis of GCSM on synthetic and real-life datasets indicates the benefit of modeling dependencies between components for reducing uncertainty and for improving performance in extrapolation tasks.

---

---

In this paper, we develop a novel Aligned-Spatial Graph Convolutional Network (ASGCN) model to learn effective features for graph classification. Our idea is to transform arbitrary-sized graphs into fixed-sized aligned grid structures, and define a new spatial graph convolution operation associated with the grid structures. We show that the proposed ASGCN model not only reduces the problems of information loss and imprecise information representation arising in existing spatially-based Graph Convolutional Network (GCN) models, but also bridges the theoretical gap between traditional Convolutional Neural Network (CNN) models and spatially-based GCN models. Moreover, the proposed ASGCN model can adaptively discriminate the importance between specified vertices during the process of spatial graph convolution, explaining the effectiveness of the proposed model. Experiments on standard graph datasets demonstrate the effectiveness of the proposed model.

---

---

This paper proposes a novel deep reinforcement learning architecture that was inspired by previous tree structured architectures which were only useable in discrete action spaces. Policy Prediction Network offers a way to improve sample complexity and performance on continuous control problems in exchange for extra computation at training time but at no cost in computation at rollout time. Our approach integrates a mix between model-free and model-based reinforcement learning. Policy Prediction Network is the first to introduce implicit model-based learning to Policy Gradient algorithms for continuous action space and is made possible via the empirically justified clipping scheme. Our experiments are focused on the MuJoCo environments so that they can be compared with similar work done in this area.

---

---

Most positive and unlabeled data is subject to selection biases. The labeled examples can, for example, be selected from the positive set because they are easier to obtain or more obviously positive. This paper investigates how learning can be enabled in this setting. We propose and theoretically analyze an empirical-risk-based method for incorporating the labeling mechanism. Additionally, we investigate under which assumptions learning is possible when the labeling mechanism is not fully understood and propose a practical method to enable this. Our empirical analysis supports the theoretical results and shows that taking into account the possibility of a selection bias, even when the labeling mechanism is unknown, improves the trained classifiers.

---

---

A gradient boosting decision tree model is a powerful machine learning method that iteratively constructs decision trees to form an additive ensemble model. The method uses the gradient of the loss function to improve the model at each iteration step.Inspired by the databaseliterature, we exploit bitset and bitslice data structures in order toimprove the run time efficiency of learning the trees. We can use thesestructures in two ways. First, they can represent the input data itself.Second, they can store the discretized gradient values used by the learningalgorithm to construct the trees in the boosting model. Using thesebit-level data structures reduces the problem of finding thebest split, which involves counting of instances and summing gradientvalues, to counting one-bits in bit strings.Modern CPUs can efficiently count one-bits using AVX2 SIMD instructions.Empirically, ourproposed improvements can result in speed-ups of 2 to up to 10 times ondatasets with a large number of categorical feature without sacrificingpredictive performance.

---

---

Motivated by various binary classification problems in structured data (e.g., graphs or other relational and algebraic structures), we investigate some algorithmic properties of closed set and half-space separation in abstract closure systems.Assuming that the underlying closure system is finite and given by the corresponding closure operator, we formulate some negative and positive complexity results for these two separation problems.In particular, we prove that deciding half-space separability in abstract closure systems is NP-complete in general.On the other hand, for the relaxed problem of maximal closed set separation we propose a simple greedy algorithm and show that it is efficient and has the best possible lower bound on the number of closure operator calls.As a second direction to overcome the negative result above, we consider Kakutani closure systems and show first that our greedy algorithm provides an algorithmic characterization of this kind of set systems.As one of the major potential application fields, we then focus on Kakutani closure systems over graphs and generalize a fundamental characterization result based on the Pasch axiom to graph structure partitioning of finite sets.Though the primary focus of this work is on the generality of the results obtained, we experimentally demonstrate the practical usefulness of our approach on vertex classification in different graph datasets.Motivated by various binary classification problems in structured data (e.g., graphs or other relational and algebraic structures), we investigate some algorithmic properties of closed set and half-space separation in abstract closure systems.Assuming that the underlying closure system is finite and given by the corresponding closure operator, we formulate some negative and positive complexity results for these two separation problems.In particular, we prove that deciding half-space separability in abstract closure systems is NP-complete in general.On the other hand, for the relaxed problem of maximal closed set separation we propose a simple greedy algorithm and show that it is efficient and has the best possible lower bound on the number of closure operator calls.As a second direction to overcome the negative result above, we consider Kakutani closure systems and show first that our greedy algorithm provides an algorithmic characterization of this kind of set systems.As one of the major potential application fields, we then focus on Kakutani closure systems over graphs and generalize a fundamental characterization result based on the Pasch axiom to graph structure partitioning of finite sets.Though the primary focus of this work is on the generality of the results obtained, we experimentally demonstrate the practical usefulness of our approach on vertex classification in different graph datasets.

---

---

Before constructing a classifier, we should examine the data to gainan understanding of the relationships between the variables, toassist with the design of the classifier. Using multi-label datarequires us to examine the association between labels: its multi-labelness. We cannotdirectly measure association between two labels, since the labels'relationships are confounded with the set of observation variables.A better approach is to fit an analytical model to a label withrespect to the observations and remaining labels, but this mightpresent false relationships due to the problem of multicollinearitybetween the observations and labels.In this article, we examine the utility of regularised logisticregression and a new form of split logistic regression for assessingthe multi-labelness of data.We find that a split analytical model usingregularisation is able to provide fewer label relationships when norelationships exist, or if the labels can be partitioned. We alsofind that if label relationships do exist, logistic regression with l_1 regularisationprovides the better measurement of multi-labelness.

---

---

This paper aims at identifying sequences of words related to specific product components in online product reviews. A reliable baseline performance for this topic classification problem is given by a Max Entropy classifier which assumes independence over subsequent topics. However, the reviews exhibit an inherent structure on the document level allowing to frame the task as sequence classification problem. Since more flexible models from the class of Conditional Random Fields were not competitive because of the limited amount of training data available, we propose using a Hidden Markov Model instead and decouple the training of transition and emission probabilities. The discriminating power of the Max Entropy approach is used for the latter.Besides outperforming both standalone methods as well as more generic models such as linear-chain Conditional Random Fields, the combined classifier is able to assign topics on sub-sentence level although labeling in the training data is only available on sentence level.

---

---

We present an approach to explain the decisions of black box models for image classification. While using the black box to label images, our explanation method exploits the latent feature space learned through an adversarial autoencoder. The proposed method first generates exemplar images in the latent feature space and learns a decision tree classifier. Then, it selects and decodes exemplars respecting local decision rules. Finally, it visualizes them in a manner that shows to the user how the exemplars can be modified to either stay within their class, or to become counter-factuals by "morphing" into another class. Since we focus on black box decision systems for image classification, the explanation obtained from the exemplars also provides a saliency map highlighting the areas of the image that contribute to its classification, and areas of the image that push it into another class. We present the results of an experimental evaluation on three datasets and two black box models. Besides providing the most useful and interpretable explanations, we show that the proposed method outperforms existing explainers in terms of fidelity, relevance, coherence, and stability.

---

---

Measuring hardness of individual instances in machine learning contributes to a deeper analysis of learning performance. This work proposes instance hardness measures for binary classification in cost-sensitive scenarios. Here cost curves are generated for each instance, defined as the loss observed for a pool of learning models for that instance along the range of cost proportions. Instance hardness is defined as the area under the cost curves and can be seen as an expected loss of difficulty along cost proportions. Different cost curves were proposed by considering common decision threshold choice methods in literature, thus providing alternative views of instance hardness.

---

---

Recently, neural networks trained as optimizers under the "learning to learn" or meta-learning framework have been shown to be effective for a broad range of optimization tasks including derivative-free black-box function optimization. Recurrent neural networks (RNNs) trained to optimize a diverse set of synthetic non-convex differentiable functions via gradient descent have been effective at optimizing derivative-free black-box functions.In this work, we propose RNN-Opt: an approach for learning RNN-based optimizers for optimizing real-parameter single-objective continuous functions under limited budget constraints.Existing approaches utilize an observed improvement based meta-learning loss function for training such models. We propose training RNN-Opt by using synthetic non-convex functions with known (approximate) optimal values by directly using discounted regret as our meta-learning loss function. We hypothesize that a regret-based loss function mimics typical testing scenarios, and would therefore lead to better optimizers compared to optimizers trained only to propose queries that improve over previous queries.Further, RNN-Opt incorporates simple yet effective enhancements during training and inference procedures to deal with the following practical challenges: (i) Unknown range of possible values for the black-box function to be optimized, and (ii) Practical and domain-knowledge based constraints on the input parameters.We demonstrate the efficacy of RNN-Opt in comparison to existing methods on several synthetic as well as standard benchmark black-box functions along with an anonymized industrial constrained optimization problem.

---

---

Reliably detecting anomalies in a given set of images is a task of high practical relevance for visual quality inspection, surveillance, or medical image analysis. Autoencoder neural networks learn to reconstruct normal images, and hence can classify those images as anomalies, where the reconstruction error exceeds some threshold. Here we analyze a fundamental problem of this approach when the training set is contaminated with a small fraction of outliers.We find that continued training of autoencoders inevitably reduces the reconstruction error of outliers, and hence degrades the anomaly detection performance. In order to counteract this effect, an adversarial autoencoder architecture is adapted, which imposes a prior distribution on the latent representation, typically placing anomalies into low likelihood-regions.Utilizing the likelihood model, potential anomalies can be identified and rejected already during training, which results in an anomaly detector that is significantly more robust to the presence of outliers during training.

---

---

Sharing knowledge between tasks is vital for efficient learning in a multi-task setting.However, most research so far has focused on the easier case where knowledge transfer is not harmful, i.e., where knowledge from one task cannot negatively impact the performance on another task.In contrast, we present an approach to multi-task deep reinforcement learning based on attention that does not require any a-priori assumptions about the relationships between tasks. Our attention network automatically groups task knowledge into sub-networks on a state level granularity. It thereby achieves positive knowledge transfer if possible, and avoids negative transfer in cases where tasks interfere. We test our algorithm against two state-of-the-art multi-task/transfer learning approaches and show comparable or superior performance while requiring fewer network parameters.

---

---

Classifiers can often output a score or a probability indicating how sure they are about the predicted class. Classifier calibration methods can map these into calibrated class probabilities, supporting cost-optimal decision making. Isotonic calibration is the standard non-parametric calibration method for binary classifiers, and it can be shown to yield the most likely monotonic calibration map on the given data, where monotonicity means that instances with higher predicted scores are more likely to be positive. Another non-parametric method is ENIR (ensemble of near-isotonic regression models) which allows for some non-monotonicity, but adds a penalty for it. We first demonstrate that these two methods tend to be over-confident and show that applying label smoothing improves calibration of both methods in more than 90% of studied cases. Unfortunately, label smoothing reduces confidence on the under-confident predictions also, and it does not reduce the raggedness of isotonic calibration. As the main contribution we propose a non-parametric Bayesian isotonic calibration method which has the flexibility of isotonic calibration to fit maps of all monotonic shapes but it adds smoothness and reduces over-confidence without requiring label smoothing. The method introduces a prior over piecewise linear monotonic calibration maps and uses a simple Monte Carlo sampling based approach to approximate the posterior mean calibration map.Our experiments demonstrate that on average the proposed method results in better calibrated probabilities than the state-of-the-art calibration methods, including isotonic calibration and ENIR.

---

---

Incorporating second order curvature information in gradient based methods have shown to improve convergence drastically despite its computational intensity. In this paper, we propose a stochastic (online) quasi-Newton method with Nesterov's accelerated gradient in both its full and limited memory forms for solving large scale non-convex optimization problems in neural networks. The performance of the proposed algorithm is evaluated in Tensorflow on benchmark classification and regression problems. The results show improved performance compared to the classical second order oBFGS and oLBFGS methods and popular first order stochastic methods such as SGD and Adam. The performance with different momentum rates and batch sizes have also been illustrated.

---

---

Uplift modeling is an approach to machine learning which allows forpredicting the net effect of an action (with respect to not taking theaction). To achieve this, the training population is divided into twoparts: the treatment group, which is subjected to the action, and thecontrol group, on which the action is not taken. Our task is toconstruct a model which will predict the difference between outcomes inthe treatment and control groups conditional on individual objects'features. When the group assignment is random, the model admits acausal interpretation. When we assume linear responses in bothgroups, the simplest way of estimating the net effect of the action onan individual is to build two separate linear ordinary least squares(OLS) regressions on the treatment and control groups and compute thedifference between their predictions. In classical linear modelsimprovements in accuracy can be achieved through the use of so calledshrinkage estimators such as the well known James-Stein estimator,which has a provably lower mean squared error than the OLSestimator. In this paper we investigate the use of shrinkageestimators in the uplift modeling problem. Unfortunately directgeneralization of the James-Stein estimator does not lead to improvedpredictions, nor does shrinking treatment and control modelsseparately. Therefore, we propose a new uplift shrinkage method whereestimators in the treatment and control groups are shrunk jointly so as tominimize the error in the predicted net effect of the action. Weprove that the proposed estimator does indeed improve on the doubleregression estimator.

---

---

Since resource-constrained devices hardly benefit from the trend towards ever-increasing neural network (NN) structures, there is growing interest in designing more hardware-friendly NNs.In this paper, we consider the training of NNs with discrete-valued weights and sign activation functions that can be implemented more efficiently in terms of inference speed, memory requirements, and power consumption.We build on the framework of probabilistic forward propagations using the local reparameterization trick, where instead of training a single set of NN weights we rather train a distribution over these weights.Using this approach, we can perform gradient-based learning by optimizing the continuous distribution parameters over discrete weights while at the same time perform backpropagation through the sign activation.In our experiments, we investigate the influence of the number of weights on the classification performance on several benchmark datasets, and we show that our method achieves state-of-the-art performance.

---

---

Identity stitching, the task of identifying and matching various online references (e.g., sessions over different devices and timespans) to the same user in real-world web services, is crucial for personalization and recommendations. However, traditional user stitching approaches, such as grouping or blocking, require quadratic pairwise comparisons between a massive number of user activities, thus posing both computational and storage challenges. Recent works, which are often application-specific, heuristically seek to reduce the amount of comparisons, but they suffer from low precision and recall. To solve the problem in an application-independent way, we take a heterogeneous network-based approach in which users (nodes) interact with content (e.g., sessions, websites), and may have attributes (e.g., location). We propose node2bits, an efficient framework that represents multi-dimensional features of node contexts with binary hashcodes.node2bits leverages feature-based temporal walks to encapsulate short- and long-term interactions between nodes in heterogeneous web networks, and adopts SimHash to obtain compact, binary representations and avoid the quadratic complexity for similarity search.Extensive experiments on large-scale real networks show that node2bits outperforms traditional techniques and existing works that generate real-valued embeddings by up to 5.16

---

---

We propose an overlapping community model based on the Affiliation Graph Model (AGM), that exhibits the pluralistic homophily property that the probability of a link between nodes increases with increasing number of shared communities. We take inspiration from the Mixed Membership Stochastic Blockmodel (MMSB), in proposing an edgewise community affiliation. This allows decoupling of community affiliations between nodes, opening the way to scalable inference. We show that our model corresponds to an AGM with soft community affiliations and develop a scalable algorithm based on a Stochastic Gradient Riemannian Langevin Dynamics(SGRLD) sampler. Empirical results show that the model can scale to network sizes that are beyond the capabilities of MCMC samplers of the standard AGM. We achieve comparable performance in terms of accuracy and run-time efficiency to scalable MMSB samplers.

---

---

Class-imbalance is an inherent characteristic of multi-label data which affects the prediction accuracy of most multi-label learning methods. One efficient strategy to deal with this problem is to employ resampling techniques before training the classifier. Existing multi-label sampling methods alleviate the (global) imbalance of multi-label datasets. However, performance degradation is mainly due to rare sub-concepts and overlapping of classes that could be analysed by looking at the local characteristics of the minority examples, rather than the imbalance of the whole dataset. We propose a new method for synthetic oversampling of multi-label data that focuses on local label distribution to generate more diverse and better labeled instances. Experimental results on 13 multi-label datasets demonstrate the effectiveness of the proposed approach in a variety of evaluation measures, particularly in the case of an ensemble of classifiers trained on repeated samples of the original data.

---

---

The development of cluster computing frameworks hasallowed practitioners to scale out various statistical estimation and machinelearning algorithms with minimal programming effort. This is especially true for machine learning problems whose objective function is nicely separable across individual data points, such as classification and regression.In contrast, statistical learning tasks involving pairs (ormore generally tuples) of data points - such as metric learning,clustering or ranking - do not lend themselves as easily todata-parallelismand in-memory computing.In this paper, we investigate how to balance between statistical performanceand computational efficiency in such distributed tuplewise statisticalproblems. We first propose a simple strategy based on occasionallyrepartitioningdata across workers between parallel computation stages, where the number ofrepartitioning steps rules the trade-off between accuracy and runtime. We thenpresent some theoretical results highlighting the benefits brought by theproposed method in terms of variance reduction, and extend our results todesigndistributed stochastic gradient descent algorithms for tuplewiseempiricalrisk minimization.Our results are supported by numerical experiments in pairwisestatistical estimation and learning on synthetic and real-world datasets.

---

---

Change point detection plays a fundamental role in many real-world applications,where the goal is to analyze and monitor the behaviour of a data stream.In this paper, we study change detection in binary streams. To this end, weuse a likelihood ratio between two models as a measure for indicatingchange. The first model is a single bernoulli variable while the second modeldivides the stored data in two segments, and models each segment with itsown bernoulli variable. Finding the optimal split can be done in O(n) time,where n is the number of entries since the last change point.This is too expensive for large n. To combat this we propose an approximationscheme that yields (1 - ϵ) approximation in O(ϵ^-1^2 n) time.The speed-up consists of several steps: First we reduce the number of possiblecandidates by adopting a known result from segmentation problems. We then showthat for fixed bernoulli parameters we can find the optimal change point inlogarithmic time. Finally, we show how to construct a candidate listof size O(ϵ^-1 n) for model parameters.We demonstrate empirically the approximation quality and the running time ofour algorithm, showing that we can gain a significant speed-up with a minimalaverage loss in optimality.

---

---

Deep neural networks achieve state-of-the-art results in various classification and synthetic data generation tasks. However, only little is known about why depth improves a model. We investigate the structure of stochastic deep neural works, also known as Deep Boltzmann Machines, to shed some light on this issue. While the best known results postulate an exponential dependence between the number of visible units and the depth of the model, we show that the required depth is upper bounded by the longest path in the underlying junction tree, which is at most linear in the number of visible units. Moreover, we show that the conditional independence structure of any categorical Deep Boltzmann Machine contains a sub-tree that allows the consistent estimation of the full joint probability mass function of all visible units. We connect our results to l_1-regularized maximum-likelihood estimation and Chow-Liu trees. Based on our theoretical findings, we present a new tractable version of Deep Boltzmann Machines, namely the Deep Boltzmann Tree (DBT). We provide a hyper-parameter-free algorithm for learning the DBT from data, and propose a new initialization method to enforce convergence to good solutions. Our findings provide some theoretical evidence for why a deep model might be beneficial. Experimental results on benchmark data show, that the DBT is a theoretical sound alternative to likelihood-free generative models.

---

---

We propose deep convolutional Gaussian processes, a deep Gaussian process architecture with convolutional structure. The model is a principled Bayesian framework for detecting hierarchical combinations of local features for image classification. We demonstrate greatly improved image classification performance compared to current convolutional Gaussian process approaches on the MNIST and CIFAR-10 datasets. In particular, we improve state-of-the-art CIFAR-10 accuracy by over 10 percentage points.

---

---

In this paper, we study the stochastic version of the one-sided full information bandit problem, where we have K arms [K] = {1, 2, ..., K}, and playing arm i would gain reward from an unknown distribution for arm iwhile obtaining reward feedback for all arms j > i.One-sided full information bandit can model the online repeated second-price auctions, where the auctioneer could select the reserved price in each round and the bidders only reveal their bids when their bids are higher than the reserved price. In this paper, we present an elimination-based algorithm to solve the problem. Our elimination based algorithm achieves distribution independent regret upper bound O(√(T· (TK))), and distribution dependent bound O(( T + K)f(Δ)), where T is the time horizon, Δ is a vector of gaps between the mean reward of arms and the mean reward of the best arm, and f(Δ) is a formula depending on the gap vector that we will specify in detail. Our algorithm has the best theoretical regret upper bound so far.We also validate our algorithm empirically against other possible alternatives.

---

---

Association rules are among the most important concepts in data mining. Rules of the form X -> Y are simple to understand, simple to act upon, yet can model important local dependencies in data. The problem is, however, that there are so many of them. Both traditional and state-of-the-art frameworkstypically yield millions of rules, rather than identifying a small set of rules that capture the most important dependencies of the data.In this paper, we define the problem of association rule mining in terms of the Minimum Description Length principle.That is, we identify the best set of rules as the one that most succinctly describes the data.We show that the resulting optimization problem does not lend itself for exact search, and hence propose Grab, a greedy heuristic to efficiently discover good sets of noise-resistant rules directly from data.Through extensive experiments we show that, unlike the state-of-the-art, Grab does reliably recover the ground truth.On real world data we show it finds reasonable numbers of rules, that upon close inspection give clear insight in the local distribution of the data.

---

---

Both complex and evolving nature of time series structure make forecasting among one of the most important and challenging tasks in time series analysis. Typical methods for forecasting are designed to model time-evolving dependencies between data observations. However, it is generally accepted that none of them is universally valid for every application. Therefore, methods for learning heterogeneous ensembles by combining a diverse set of forecasts together appear as a promising solution to tackle this task. Hitherto, in classical ML literature, ensemble techniques such as stacking, cascading and voting are mostly restricted to operate in a static manner. To deal with changes in the relative performance of models as well as changes in the data distribution, we propose a drift-aware meta-learning approach for adaptively selecting and combining forecasting models. Our assumption is that different forecasting models have different areas of expertise and a varying relative performance. Our method ensures dynamic selection of initial ensemble base models candidates through a performance drift detection mechanism. Since diversity is a fundamental component in ensemble methods, we propose a second stage selection with clustering that is computed after each drift detection. Predictions of final selected models are combined into a single prediction. An exhaustive empirical testing of the method was performed, evaluating both generalization error and scalability of the approach using time series from several real world domains. Empirical results show the competitiveness of the method in comparison to state-of-the-art approaches for combining forecasters.

---

---

Signed networks contain both positive and negative kinds of interactions like friendship and enmity. The task of node classification in non-signed graphs has proven to be beneficial in many real world applications, yet extensions to signed networks remain largely unexplored. In this paper we introduce the first analysis of node classification in signed social networks via diffuse interface methods based on the Ginzburg-Landau functional together with different extensions of the graph Laplacian to signed networks. We show that blending the information from both positive and negative interactions leads to performance improvement in real signed social networks, consistently outperforming the current state of the art.

---

---

Existing cluster validity indices often possess a similar bias as theclustering algorithm they were introduced for, e.g. to determine theoptimal number of clusters. We suggest an efficient and holisticassessment of the structure discovery capabilities of clusteringalgorithms based on three criteria. We determine the robustness orstability of cluster assignments and interpret it as the confidence ofthe clustering algorithm in its result. This information is then usedto label the data and evaluate the consistency of thestability-assessment with the notion of a cluster as an area of denseand separated data. The resulting criteria of stability, structure andconsistency provide interpretable means to judge the capabilities ofclustering algorithms without the typical biases of prominent indices,including the judgment of a clustering tendency.

---

---

The present-day accessibility of technology enables easy logging of both sensor values and event logs over extended periods. In this context, detecting abnormal segments in time series data has become an important data mining task. Existing work on anomaly detection focuses either on continuous time series or discrete event logs and not on the combination.However, in many practical applications, the patterns extracted from the event log can reveal contextual and operational conditions of a device that must be taken into account when predicting anomalies in the continuous time series.This paper proposes an anomaly detection method that can handle mixed-type time series. The method leverages frequent pattern mining techniques to construct an embedding of mixed-type time series on which an isolation forest is trained. Experiments on several real-world univariate and multivariate time series, as well as a synthetic mixed-type time series, show that our anomaly detection algorithm outperforms state-of-the-art anomaly detection techniques such as MatrixProfile, Pav, Mifpod and Fpof.

---

---

With Sobolev Training, neural networks are trained to fit target output values as well as target derivatives with respect to the inputs. This leads to better generalization and fewer required training examples for certain problems. In this paper, we present a training pipeline that enables Sobolev Training for regression problems where target derivatives are not directly available. Thus, we propose to use a least-squares estimate of the target derivatives based on function values of neighboring training samples. We show for a variety of black-box function regression tasks that our training pipeline achieves smaller test errors compared to the traditional training method. Since our method has no additional requirements on the data collection process, it has great potential to improve the results for various regression tasks.

---

---

Unsupervised feature selection is mostly assessed along a supervised learning setting, depending on whether the selected features efficiently permit to predict the (unknown) target variable. Another setting is proposed in this paper: the selected features aim to efficiently recover the whole dataset. The proposed algorithm, called AgnoS, combines an AutoEncoder with structural regularizations to sidestep the combinatorial optimization problem at the core of feature selection. The extensive experimental validation of AgnoS on the scikit-feature benchmark suite demonstrates its ability compared to the state of the art, both in terms of supervised learning and data compression.

---

---

We propose a Bayesian information-geometric approach to the exploration-exploitation trade-off in stochastic multi-armed bandits. The uncertainty on reward generation and belief is represented using the manifold of joint distributions of rewards and beliefs. Accumulated information is summarised by the barycentre of joint distributions, the pseudobelief-reward. While the pseudobelief-reward facilitates information accumulation through exploration, another mechanism is needed to increase exploitation by gradually focusing on higher rewards, the pseudobelief-focal-reward. Our resulting algorithm, BelMan, alternates between projection of the pseudobelief-focal-reward onto belief-reward distributions to choose the arm to play, and projection of the updated belief-reward distributions onto the pseudobelief-focal-reward. We theoretically prove BelMan to be asymptotically optimal and to incur a sublinear regret growth. We instantiate BelMan to stochastic bandits with Bernoulli and exponential rewards, and to a real-life application of scheduling queueing bandits. Comparative evaluation with the state of the art shows that BelMan is not only competitive for Bernoulli bandits but in many cases also outperforms other approaches for exponential and queueing bandits.

---

---

We consider network sparsification as an L_0-norm regularized binary optimization problem, where each unit of a neural network (e.g., weight, neuron, or channel, etc.) is attached with a stochastic binary gate, whose parameters are jointly optimized with original network parameters. The Augment-Reinforce-Merge (ARM), a recently proposed unbiased gradient estimator, is investigated for this binary optimization problem. Compared to the hard concrete gradient estimator from Louizos et al., ARM demonstrates superior performance of pruning network architectures while retaining almost the same accuracies of baseline methods. Similar to the hard concrete estimator, ARM also enables conditional computation during model training but with improved effectiveness due to the exact binary stochasticity. Thanks to the flexibility of ARM, many smooth or non-smooth parametric functions, such as scaled sigmoid or hard sigmoid, can be used to parameterize this binary optimization problem and the unbiasness of the ARM estimator is retained, while the hard concrete estimator has to rely on the hard sigmoid function to achieve conditional computation and thus accelerated training. Extensive experiments on multiple public datasets demonstrate state-of-the-art pruning rates with almost the same accuracies of baseline methods. The resulting algorithm L_0-ARM sparsifies the Wide-ResNet models on CIFAR-10 and CIFAR-100 while the hard concrete estimator cannot.

---

---

In neural network optimization, the learning rate of the gradient descent strongly affects performance. This prevents reliable out-of-the-box training of a model on a new problem. We propose theAll Learning Rates At Once (Alrao) algorithm fordeep learning architectures:each neuron or unit in the network gets its own learning rate, randomly sampled atstartupfrom a distribution spanning several orders of magnitude. The network becomes a mixture of slow and fast learning units. Surprisingly, Alrao performs close to SGD with an optimally tuned learning rate, for various tasks and network architectures. In our experiments, all Alrao runs were able to learn well without any tuning.

---

---

A triangle is an important building block of social networks, so the study of triangleformation in a network is critical for better understanding of the dynamics of suchnetworks. Existing works in this area mainly focus on triangle counting, or generating synthetic networks by matching the prevalence of trianglesin real-life networks. While these efforts increase our understanding of triangle's role in a network, they have limited practical utility. In this work we undertake aninteresting problem relating to triangle formation in a network, which is, to predict the time by which the third link of a triangle appears in a network. Since the thirdlink completes a triangle, we name this task as Triangle CompletionTime Prediction (TCTP). Solution to TCTP problem is valuable forreal-life link recommendation in social/e-commerce networks, also it provides vital information for dynamic network analysis and community generation study.An efficient and robust framework (GraNiTE) is proposed for solving the TCTP problem.GraNiTE uses neural networks based approach for learning a representation vector of a trianglecompleting edge, which is a concatenation of two representation vectors: first one is learnt from graphlet based local topology around that edge and the second one islearnt from time-preserving embedding of the constituting vertices of that edge. A comparison of the proposed solution with several baselinemethods shows that the mean absolute error (MAE) of the proposed method is at least one-forth of that of the best baseline method.

---

---

Most convex and nonconvex clustering algorithms come with one crucial parameter: the k in k-means. To this day, there is not one generally accepted way to accurately determine this parameter. Popular methods are simple yet theoretically unfounded, such as searching for an elbow in the curve of a given cost measure. In contrast, statistically founded methods often make strict assumptions over the data distribution or come with their own optimization scheme for the clustering objective. This limits either the set of applicable datasets or clustering algorithms. In this paper, we strive to determine the number of clusters by answering a simple question: given two clusters, is it likely that they jointly stem from a single distribution? To this end, we propose a bound on the probability that two clusters originate from the distribution of the unified cluster, specified only by the sample mean and variance. Our method is applicable as a simple wrapper to the result of any clustering method minimizing the objective of k-means, which includes Gaussian mixtures and Spectral Clustering. We focus in our experimental evaluation on an application for nonconvex clustering and demonstrate the suitability of our theoretical results. Our SpecialK clustering algorithm automatically determines the appropriate value for k, without requiring any data transformation or projection, and without assumptions on the data distribution. Additionally, it is capable to decide that the data consists of only a single cluster, which many existing algorithms cannot.

---

---

In this paper, we propose a theoretically founded sequentialstrategy for training large-scale Recommender Systems (RS) over implicitfeedback mainly in the form of clicks. The proposed approach consistsin minimizing pairwise ranking loss over blocks of consecutive itemsconstituted by a sequence of non-clicked items followed by a clicked onefor each user. Parameter updates are discarded if for a given user thenumber of sequential blocks is below or above some given thresholdsestimated over the distribution of the number of blocks in the training set.This is to prevent from updating the parameters for an abnormally highnumber of clicks over some targeted items, mainly due to bots; or very fewuser interactions. Both scenarios affect the decision of RS and imply a shiftover the distribution of items that are shown to the users. We provide aproof of convergence of the algorithm to the minimizer of the ranking loss,in the case where the latter is convex. Furthermore, experimental resultson five large-scale collections demonstrate the efficiency of the proposedalgorithm concerning the state-of-the-art approaches, both regardingdifferent ranking measures and computation time.

---

---

Recently, generating adversarial examples has become an important means of measuring robustness of a deep learning model. Adversarial examples help us identify the susceptibilities of the model and further counter those vulnerabilities by applying adversarial training techniques. In natural language domain, small perturbations in the form of misspellings or paraphrases can drastically change the semantics of the text. We propose a reinforcement learning based approach towards generating adversarial examples in black-box settings. We demonstrate that our method is able to fool well-trained models for (a) IMDB sentiment classification task and (b) AG's news corpus news categorization task with significantly high success rates. We find that the adversarial examples generated are semantics-preserving perturbations to the original text.

---

---

We propose FastPoint, a novel multivariate point process that enables fast and accurate learning and inference. FastPoint uses deep recurrent neural networks to capture complex temporal dependency patterns among different marks, while self-excitation dynamics within each mark are modeled with Hawkes processes. This results in substantially more efficient learning and scales to millions of correlated marks with superior predictive accuracy. Our construction also allows for efficient and parallel sequential Monte Carlo sampling for fast predictive inference. FastPoint outperforms baseline methods in prediction tasks on synthetic and real-world high-dimensional event data at a small fraction of the computational cost.

---

---

Conditional Generative Models are now acknowledged an essential tool in Machine Learning. This paper focuses on their control. While many approaches aim at disentangling the data through the coordinate-wise control of their latent representations, another direction is explored in this paper. The proposed CompVAE handles data with a natural multi-ensemblist structure (i.e. that can naturally be decomposed into elements). Derived from Bayesian variational principles, CompVAE learns a latent representation leveraging both observational and symbolic information. A first contribution of the approach is that this latent representation supports a compositional generative model, amenable to multi-ensemblist operations (addition or subtraction of elements in the composition). This compositional ability is enabled by the invariance and generality of the whole framework w.r.t. respectively, the order and number of the elements. The second contribution of the paper is a proof of concept on synthetic 1D and 2D problems, demonstrating the efficiency of the proposed approach.

---

---

Can we automatically design a Convolutional Network (ConvNet) with thehighest image classification accuracy under the latency constraint of a mobile device? Neural architecture search (NAS) has revolutionizedthe design of hardware-efficient ConvNets by automating this process.However, the NAS problem remains challenging due to the combinatorially large design space, causing a significant searching time (at least 200 GPU-hours). To alleviate this complexity, we propose Single-Path NAS, a novel differentiable NAS method for designing hardware-efficient ConvNets in less than 4 hours. Our contributions are as follows: 1. Single-path search space: Compared to previous differentiable NAS methods, Single-Path NAS uses one single-path over-parameterized ConvNet to encode all architectural decisions with shared convolutionalkernel parameters, hence drastically decreasing the number oftrainable parameters and the search cost down to few epochs. 2. Hardware-efficient ImageNet classification: Single-Path NAS achieves 74.96

---

---

Bayesian global-local shrinkage estimation with the generalized horseshoe prior represents the state-of-the-art for Gaussian regression models. The extension to non-Gaussian data, such as binary or Student-t regression, is usually done by exploiting a scale-mixture-of-normals approach. However, many standard distributions, such as the gamma and the Poisson, do not admit such a representation. We contribute two extensions to global-local shrinkage methodology. The first is an adaption of recent auxiliary gradient based-sampling schemes to the global-local shrinkage framework, which yields simple algorithms for sampling from generalized linear models. We also introduce two new samplers for the hyperparameters in the generalized horseshoe model, one based on an inverse-gamma mixture of inverse-gamma distributions, and the second a rejection sampler. Results show that these new samplers are highly competitive with the no U-turn sampler for small numbers of predictors, and potentially perform better for larger numbers of predictors. Results for hyperparameter sampling show our new inverse-gamma inverse-gamma based sampling scheme outperforms the standard sampler based on a gamma mixture of gamma distributions.

---

---

Many real-world security scenarios can be modeled via a game-theoretic framework known as a security game in which there is a defender trying to protect potential targets from an attacker. Recent work in security games has shown that deceptive signaling by the defender can convince an attacker to withdraw his attack. For instance, a warning message to commuters indicating speed enforcement is in progress ahead might lead to them driving more slowly, even if it turns out no enforcement is in progress. However, the results of this work are limited by the unrealistic assumption that the attackers will behave with perfect rationality, meaning they always choose an action that gives them the best expected reward. We address the problem of training boundedly rational (human) attackers to comply with signals via repeated interaction with signaling without incurring a loss to the defender, and offer the four following contributions: (i) We learn new decision tree and neural network-based models of attacker compliance with signaling. (ii) Based on these machine learning models of a boundedly rational attacker's response to signaling, we develop a theory of signaling in the Goldilocks zone, a balance of signaling and deception that increases attacker compliance and improves defender utility. (iii) We present game-theoretic algorithms to solve for signaling schemes based on the learned models of attacker compliance with signaling. (iv) We conduct extensive human subject experiments using an online game. The game simulates the scenario of an inside attacker trying to steal sensitive information from company computers, and results show that our algorithms based on learned models of attacker behavior lead to better attacker compliance and improved defender utility compared to the state-of-the-art algorithm for rational attackers with signaling.

---

---

Explaining the results of Machine learning algorithms is crucial given the rapid growth and potential applicability of these methods in critical domains including healthcare, defense, autonomous driving, etc. In this paper, we address this problem in the context of Markov Logic Networks (MLNs) which are highly expressive statistical relational models that combine first-order logic with probabilistic graphical models. MLNs in general are known to be interpretable models, i.e., MLNs can be understood more easily by humans as compared to models learned by approaches such as deep learning. However, at the same time, it is not straightforward to obtain human-understandable explanations specific to an observed inference result (e.g. marginal probability estimate). This is because, the MLN provides a lifted interpretation, one that generalizes to all possible worlds/instantiations, which are not query/evidence specific. In this paper, we extract grounded-explanations, i.e., explanations defined w.r.t specific inference queries and observed evidence. We extract these explanations from importance weights defined over the MLN formulas that encode the contribution of formulas towards the final inference results. We validate our approach in real world problems related to analyzing reviews from Yelp, and show through user-studies that our explanations are richer than state-of-the-art non-relational explainers such as LIME.

---