Optimization algorithms for training deep models not only affects the convergence rate and stability of the training process, but are also highly related to the generalization performance of the models. While adaptive algorithms, such as Adam and RMSprop, have shown better optimization performance than stochastic gradient descent (SGD) in many scenarios, they often lead to worse generalization performance than SGD, when used for training deep neural networks (DNNs). In this work, we identify two problems of Adam that may degrade the generalization performance. As a solution, we propose the normalized direction-preserving Adam (ND-Adam) algorithm, which combines the best of both worlds, i.e., the good optimization performance of Adam, and the good generalization performance of SGD. In addition, we further improve the generalization performance in classification tasks, by using batch-normalized softmax. This study suggests the need for more precise control over the training process of DNNs.

To the best of our knowledge, this paper presents the first large-scale study that tests whether network categories (e.g., social networks vs. web graphs) are distinguishable from one another (using both categories of real-world networks and synthetic graphs). A classification accuracy of $94.2\%$ was achieved using a random forest classifier with both real and synthetic networks. This work makes two important findings. First, real-world networks from various domains have distinct structural properties that allow us to predict with high accuracy the category of an arbitrary network. Second, classifying synthetic networks is trivial as our models can easily distinguish between synthetic graphs and the real-world networks they are supposed to model.

Multi-scanner Antivirus systems provide insightful information on the nature of a suspect application; however there is often a lack of consensus and consistency between different Anti-Virus engines. In this article, we analyze more than 250 thousand malware signatures generated by 61 different Anti-Virus engines after analyzing 82 thousand different Android malware applications. We identify 41 different malware classes grouped into three major categories, namely Adware, Harmful Threats and Unknown or Generic signatures. We further investigate the relationships between such 41 classes using community detection algorithms from graph theory to identify similarities between them; and we finally propose a Structure Equation Model to identify which Anti-Virus engines are more powerful at detecting each macro-category. As an application, we show how such models can help in identifying whether Unknown malware applications are more likely to be of Harmful or Adware type.

We propose a new method for embedding graphs while preserving directed edge information. Learning such continuous-space vector representations (or embeddings) of nodes in a graph is an important first step for using network information (from social networks, user-item graphs, knowledge bases, etc.) in many machine learning tasks. Unlike previous work, we (1) explicitly model an edge as a function of node embeddings, and we (2) propose a novel objective, the "graph likelihood", which contrasts information from sampled random walks with non-existent edges. Individually, both of these contributions improve the learned representations, especially when there are memory constraints on the total size of the embeddings. When combined, our contributions enable us to significantly improve the state-of-the-art by learning more concise representations that better preserve the graph structure. We evaluate our method on a variety of link-prediction task including social networks, collaboration networks, and protein interactions, showing that our proposed method learn representations with error reductions of up to 76% and 55%, on directed and undirected graphs. In addition, we show that the representations learned by our method are quite space efficient, producing embeddings which have higher structure-preserving accuracy but are 10 times smaller.

The current work characterizes the users of a VoD streaming space through user-personas based on a tenure timeline and temporal behavioral features in the absence of explicit user profiles. A combination of tenure timeline and temporal characteristics caters to business needs of understanding the evolution and phases of user behavior as their accounts age. The personas constructed in this work successfully represent both dominant and niche characterizations while providing insightful maturation of user behavior in the system. The two major highlights of our personas are demonstration of stability along tenure timelines on a population level, while exhibiting interesting migrations between labels on an individual granularity and clear interpretability of user labels. Finally, we show a trade-off between an indispensable trio of guarantees, relevance-scalability-interpretability by using summary information from personas in a CTR (Click through rate) predictive model. The proposed method of uncovering latent personas, consequent insights from these and application of information from personas to predictive models are broadly applicable to other streaming based products.

Approximate Markov chain Monte Carlo (MCMC) offers the promise of more rapid sampling at the cost of more biased inference. Since standard MCMC diagnostics fail to detect these biases, researchers have developed computable Stein discrepancy measures that provably determine the convergence of a sample to its target distribution. This approach was recently combined with the theory of reproducing kernels to define a closed-form kernel Stein discrepancy (KSD) computable by summing kernel evaluations across pairs of sample points. We develop a theory of weak convergence for KSDs based on Stein's method, demonstrate that commonly used KSDs fail to detect non-convergence even for Gaussian targets, and show that kernels with slowly decaying tails provably determine convergence for a large class of target distributions. The resulting convergence-determining KSDs are suitable for comparing biased, exact, and deterministic sample sequences and simpler to compute and parallelize than alternative Stein discrepancies. We use our tools to compare biased samplers, select sampler hyperparameters, and improve upon existing KSD approaches to one-sample hypothesis testing and sample quality improvement.

Statistical methods for network data have received a lot of attention because of the wideranging applications of network analysis. There is now a large body of work on methods and models for networks, including the stochastic block model (SBM) [Holland et al., 1983], the degree-corrected stochastic block model (DCSBM) [Karrer and Newman, 2011], and the latent space model [Hoff et al., 2002], to name a few. While this gives the practitioner plenty of choices, there is a lot less work on the crucial question of how to select the best model for the data, as well as how to choose tuning parameters for the selected model, which is often necessary in order to fit it. In some specific problems, progress has been made recently, for instance, in the much-studied problem of community detection. Community detection is the problem of clustering network nodes into groups, and most of the methods proposed over the last twenty years or so require the number of communities K as input.

We develop and apply an approach for analyzing multi-curve data where each curve is driven by a latent state process. The state at any particular point determines a smooth function, forcing the individual curve to switch from one function to another. Thus each curve follows what we call a switching nonparametric regression model. We develop an EM algorithm to estimate the model parameters. We also obtain standard errors for the parameter estimates of the state process. We consider several types of state processes: independent and identically distributed, independent but depending on a covariate and Markov. Simulation studies show the frequentist properties of our estimates. We apply our methods to a data set of a building's power usage.

However, these are not very elegant approaches to solving multiclass problems. A better alternative is provided by the construction of multiclass SVMs, where we build a two-class classifier over a feature vector derived from the pair consisting of the input features and the class of the datum. At test time, the classifier chooses the class . The margin during training is the gap between this value for the correct class and for the nearest other class, and so the quadratic program formulation will require that . This general method can be extended to give a multiclass formulation of various kinds of linear classifiers.

Five years ago, in 2012, I decided to experiment in improving my teaching by creating a flipped classroom (and semi-MOOC) for my course "Business Analytics Using Data Mining" (BADM) at the Indian School of Business. I initially designed the course at University of Maryland's Smith School of Business in 2005 and taught it until 2010. When I joined ISB in 2011 I started teaching multiple sections of BADM (which was started by Ravi Bapna in 2006), and the course was fast growing in popularity. Repeating the same lectures in multiple course sections made me realize it was time for scale! I therefore created 30 videos, covering various supervised methods (k-NN, linear and logistic regression, trees, naive Bayes, etc.) and unsupervised methods (principal components analysis, clustering, association rules), as well as important principles such as performance evaluation, the notion of a holdout set, and more.