Advances in sensor technology have enabled the collection of large-scale datasets. Such datasets can be extremely noisy and often contain a significant amount of outliers that result from sensor malfunction or human operation faults. In order to utilize such data for real-world applications, it is critical to detect outliers so that models built from these datasets will not be skewed by outliers. In this paper, we propose a new outlier detection method that utilizes the correlations in the data (e.g., taxi trip distance vs. trip time). Different from existing outlier detection methods, we build a robust regression model that explicitly models the outliers and detects outliers simultaneously with the model fitting. We validate our approach on real-world datasets against methods specifically designed for each dataset as well as the state of the art outlier detectors. Our outlier detection method achieves better performances, demonstrating the robustness and generality of our method. Last, we report interesting case studies on some outliers that result from atypical events.

We present online prediction methods for time series that let us explicitly handle nonstationary artifacts (e.g. trend and seasonality) present in most real time series. Specifically, we show that applying appropriate transformations to such time series before prediction can lead to improved theoretical and empirical prediction performance. Moreover, since these transformations are usually unknown, we employ the learning with experts setting to develop a fully online method (NonSTOP-NonSTationary Online Prediction) for predicting nonstationary time series. This framework allows for seasonality and/or other trends in univariate time series and cointegration in multivariate time series. Our algorithms and regret analysis subsume recent related work while significantly expanding the applicability of such methods. For all the methods, we provide sub-linear regret bounds using relaxed assumptions. The theoretical guarantees do not fully capture the benefits of the transformations, thus we provide a data-dependent analysis of the follow-the-leader algorithm that provides insight into the success of using such transformations. We support all of our results with experiments on simulated and real data.

We describe our first-place solution to the Animal Behavior Challenge (ABC 2018) on predicting gender of bird from its GPS trajectory. The task consisted in predicting the gender of shearwater based on how they navigate themselves across a big ocean. The trajectories are collected from GPS loggers attached on shearwaters' body, and represented as a variable-length sequence of GPS points (latitude and longitude), and associated meta-information, such as the sun azimuth, the sun elevation, the daytime, the elapsed time on each GPS location after starting the trip, the local time (date is trimmed), and the indicator of the day starting the from the trip. We used ensemble of several variants of Gradient Boosting Classifier along with Gaussian Process Classifier and Support Vector Classifier after extensive feature engineering and we ranked first out of 74 registered teams. The variants of Gradient Boosting Classifier we tried are CatBoost (Developed by Yandex), LightGBM (Developed by Microsoft), XGBoost (Developed by Distributed Machine Learning Community). Our approach could easily be adapted to other applications in which the goal is to predict a classification output from a variable-length sequence.

Abstract--As opposed to manual feature engineering which is tedious and difficult to scale, network representation learning has attracted a surge of research interests as it automates the process of feature learning on graphs. The learned lowdimensional node vector representation is generalizable and eases the knowledge discovery process on graphs by enabling various off-the-shelf machine learning tools to be directly applied. Recent research has shown that the past decade of network embedding approaches either explicitly factorize a carefully designed matrix to obtain the low-dimensional node vector representation or are closely related to implicit matrix factorization, with the fundamental assumption that the factorized node connectivity matrix is low-rank. Nonetheless, the global low-rank assumption does not necessarily hold especially when the factorized matrix encodes complex node interactions, and the resultant single low-rank embedding matrix is insufficient to capture all the observed connectivity patterns. In this regard, we propose a novel multilevel network embedding framework BoostNE, which can learn multiple network embedding representations of different granularity from coarse to fine without imposing the prevalent global low-rank assumption. The proposed BoostNE method is also in line with the successful gradient boosting method in ensemble learning as multiple weak embeddings lead to a stronger and more effective one. We assess the effectiveness of the proposed BoostNE framework by comparing it with existing state-of-the-art network embedding methods on various datasets, and the experimental results corroborate the superiority of the proposed BoostNE network embedding framework. Learning meaningful and discriminative representations of nodes in a network is essential for various network analytical tasks as it avoids the laborious manual feature engineering process.

A co-clustering model for continuous data that relaxes the identically distributed assumption within blocks of traditional co-clustering is presented. The proposed model, although allowing more flexibility, still maintains the very high degree of parsimony achieved by traditional co-clustering. A stochastic EM algorithm along with a Gibbs sampler is used for parameter estimation and an ICL criterion is used for model selection. Simulated and real datasets are used for illustration and comparison with traditional co-clustering. Clustering is the process of finding and analyzing underlying group structures in heterogenous data.

Feature selection methods are widely used in order to solve the 'curse of dimensionality' problem. Many proposed feature selection frameworks, treat all data points equally; neglecting their different representation power and importance. In this paper, we propose an unsupervised hypergraph feature selection method via a novel point-weighting framework and low-rank representation that captures the importance of different data points. We introduce a novel soft hypergraph with low complexity to model data. Then, we formulate the feature selection as an optimization problem to preserve local relationships and also global structure of data. Our approach for global structure preservation helps the framework overcome the problem of unavailability of data labels in unsupervised learning. The proposed feature selection method treats with different data points based on their importance in defining data structure and representation power. Moreover, since the robustness of feature selection methods against noise and outlier is of great importance, we adopt low-rank representation in our model. Also, we provide an efficient algorithm to solve the proposed optimization problem. The computational cost of the proposed algorithm is lower than many state-of-the-art methods which is of high importance in feature selection tasks. We conducted comprehensive experiments with various evaluation methods on different benchmark data sets. These experiments indicate significant improvement, compared with state-of-the-art feature selection methods.

Heterogeneous treatment effects are the center of gravity in many modern causal inference applications. In this paper, we investigate the estimation and inference of heterogeneous treatment effects with precision in a general nonparametric setting. To this end, we enhance the classical $k$-nearest neighbor method with a simple algorithm, extend it to a distributional setting, and suggest the two-scale distributional nearest neighbors (DNN) estimator with reduced finite-sample bias. Our recipe is first to subsample the data and average the 1-nearest neighbor estimators from each subsample. With appropriately chosen subsampling scale, the resulting DNN estimator is proved to be asymptotically unbiased and normal under mild regularity conditions. We then proceed with combining DNN estimators with different subsampling scales to further reduce bias. Our theoretical results on the advantages of the new two-scale DNN framework are well supported by several Monte Carlo simulations. The newly suggested method is also applied to a real-life data set to study the heterogeneity of treatment effects of smoking on children's birth weights across mothers' ages.

With the rising influence of machine learning algorithms on many important aspects of our daily lives, there are growing concerns that biases inherent in data can lead the behavior of these algorithms to discriminate against certain populations [1, 2, 4, 6, 8, 28, 29, 15]. In recent years, substantial research effort has been devoted to the development of mathematical definitions of bias, or its opposite, fairness, in algorithms and in data [15, 18, 26, 23, 19, 32]. In this work, we focus on the fairness scenario where there are multiple protected attributes that we aim to ensure fairness for, and which may potentially overlap with each other, such as gender, race, and sexual orientation. Our guiding principle is intersectionality, the core theoretical framework underlying the thirdwave feminist movement [13]. The principle of intersectionality states that racism, sexism, and other social systems which harm marginalized groups are interlocking in their effects, such that the lived experience of, e.g., black women, is very different than that of, e.g., white women. Intersectionality was defined by Kimberlé Crenshaw in the 1980's [13] and popularized in the 1990's, e.g. by Patricia Hill Collins [10], although the ideas are much older [11, 35]. In the context of machine learning and fairness, intersectionality was recently considered by [7], who studied the impact of the intersection of gender and skin color on computer vision performance, and by [23, 19], who aimed to protect certain subgroups in order to prevent "fairness gerrymandering."

Embedding-based methods for knowledge base completion (KBC) learn representations of entities and relations in a vector space, along with the scoring function to estimate the likelihood of relations between entities. The learnable class of scoring functions is designed to be expressive enough to cover a variety of real-world relations, but this expressive comes at the cost of an increased number of parameters. In particular, parameters in these methods are superfluous for relations that are either symmetric or antisymmetric. To mitigate this problem, we propose a new L1 regularizer for Complex Embeddings, which is one of the state-of-the-art embedding-based methods for KBC. This regularizer promotes symmetry or antisymmetry of the scoring function on a relation-by-relation basis, in accordance with the observed data. Our empirical evaluation shows that the proposed method outperforms the original Complex Embeddings and other baseline methods on the FB15k dataset.

Real-world networks usually have community structure, that is, nodes are grouped into densely connected communities. Community detection is one of the most popular and best-studied research topics in network science and has attracted attention in many different fields, including computer science, statistics, social sciences, among others. Numerous approaches for community detection have been proposed in literature, from ad-hoc algorithms to systematic model-based approaches. The large number of available methods leads to a fundamental question: whether a certain method can provide consistent estimates of community labels. The stochastic blockmodel (SBM) and its variants provide a convenient framework for the study of such problems. This article is a survey on the recent theoretical advances of community detection. The authors review a number of community detection methods and their theoretical properties, including graph cut methods, profile likelihoods, the pseudo-likelihood method, the variational method, belief propagation, spectral clustering, and semidefinite relaxations of the SBM. The authors also briefly discuss other research topics in community detection such as robust community detection, community detection with nodal covariates and model selection, as well as suggest a few possible directions for future research.