Decentralized training has been actively studied in recent years. Although a wide variety of methods have been proposed, yet the decentralized momentum SGD method is still underexplored. In this paper, we propose a novel periodic decentralized momentum SGD method, which employs the momentum schema and periodic communication for decentralized training. With these two strategies, as well as the topology of the decentralized training system, the theoretical convergence analysis of our proposed method is difficult. We address this challenging problem and provide the condition under which our proposed method can achieve the linear speedup regarding the number of workers. Furthermore, we also introduce a communication-efficient variant to reduce the communication cost in each communication round. The condition for achieving the linear speedup is also provided for this variant. To the best of our knowledge, these two methods are all the first ones achieving these theoretical results in their corresponding domain. We conduct extensive experiments to verify the performance of our proposed two methods, and both of them have shown superior performance over existing methods.

Lyme disease is a rapidly growing illness that remains poorly understood within the medical community. Critical questions about when and why patients respond to treatment or stay ill, what kinds of treatments are effective, and even how to properly diagnose the disease remain largely unanswered. We investigate these questions by applying machine learning techniques to a large scale Lyme disease patient registry, MyLymeData, developed by the nonprofit LymeDisease.org. We apply various machine learning methods in order to measure the effect of individual features in predicting participants' answers to the Global Rating of Change (GROC) survey questions that assess the self-reported degree to which their condition improved, worsened, or remained unchanged following antibiotic treatment. We use basic linear regression, support vector machines, neural networks, entropy-based decision tree models, and $k$-nearest neighbors approaches. We first analyze the general performance of the model and then identify the most important features for predicting participant answers to GROC. After we identify the "key" features, we separate them from the dataset and demonstrate the effectiveness of these features at identifying GROC. In doing so, we highlight possible directions for future study both mathematically and clinically.

Deep classifiers have achieved great success in visual recognition. However, real-world data is long-tailed by nature, leading to the mismatch between training and testing distributions. In this report, we introduce Balanced Activation (Balanced Softmax and Balanced Sigmoid), an elegant unbiased, and simple extension of Sigmoid and Softmax activation function, to accommodate the label distribution shift between training and testing in object detection. We derive the generalization bound for multiclass Softmax regression and show our loss minimizes the bound. In our experiments, we demonstrate that Balanced Activation generally provides ~3% gain in terms of mAP on LVIS-1.0 and outperforms the current state-of-the-art methods without introducing any extra parameters.

Ethical concerns regarding artificial intelligence has led to increased self-scrutiny from the machine learning community. Algorithmic fairness has proved to be a challenging research area. For one, a broadly appealing definition of fairness has long eluded philosophers, social scientists, and, more recently, the machine learning community. Currently, finding mathematical formulations of fairness is an active area of research in algorithmic fairness [Dwork et al., 2012, Chouldechova, 2016, Joseph et al., 2016]. While it is easy to agree on what is unfair, it is much harder to agree on what, exactly, is fair. For instance, ProPublica's eponymous article on machine bias [Angwin et al., 2016] uncovered prejudice against African-Americans in COMPAS (Correctional Offender Management Profiling for Alternative Sanctions), a recidivism prediction tool developed by Northpointe. But it turns out if different fairness criteria are used than those used in the ProPublica investigation, COMPAS can be more favourably viewed [Dieterich et al., 2016, Corbett-Davies et al., 2017]. This type of dissent is unavoidable as several works have shown that certain fairness criteria cannot be simultaneously satisfied [Hardt et al., 2016, Kleinberg, 2018]. Leaving aside for now the debate over the correct definition of fairness, most works in algorithmic fairness are quite straightforward once a fairness measure is settled on.

Data deduplication is the task of detecting records in a database that correspond to the same real-world entity. Our goal is to develop a procedure that samples uniformly from the set of entities present in the database in the presence of duplicates. We accomplish this by a two-stage process. In the first step, we estimate the frequencies of all the entities in the database. In the second step, we use rejection sampling to obtain a (approximately) uniform sample from the set of entities. However, efficiently estimating the frequency of all the entities is a non-trivial task and not attainable in the general case. Hence, we consider various natural properties of the data under which such frequency estimation (and consequently uniform sampling) is possible. Under each of those assumptions, we provide sampling algorithms and give proofs of the complexity (both statistical and computational) of our approach. We complement our study by conducting extensive experiments on both real and synthetic datasets.

This paper describes the approach proposed by the D2KLab team for the 2020 RecSys Challenge on the task of predicting user engagement facing tweets. This approach relies on two distinct stages. First, relevant features are learned from the challenge dataset. These features are heterogeneous and are the results of different learning modules such as handcrafted features, knowledge graph embeddings, sentiment analysis features and BERT word embeddings. Second, these features are provided in input to an ensemble system based on XGBoost. This approach, only trained on a subset of the entire challenge dataset, ranked 22 in the final leaderboard.

Stress minimization is among the best studied force-directed graph layout methods because it reliably yields high-quality layouts. It thus comes as a surprise that a novel approach based on stochastic gradient descent (Zheng, Pawar and Goodman, TVCG 2019) is claimed to improve on state-of-the-art approaches based on majorization. We present experimental evidence that the new approach does not actually yield better layouts, but that it is still to be preferred because it is simpler and robust against poor initialization.

Graph learning has emerged as a promising technique for multi-view clustering with its ability to learn a unified and robust graph from multiple views. However, existing graph learning methods mostly focus on the multi-view consistency issue, yet often neglect the inconsistency across multiple views, which makes them vulnerable to possibly low-quality or noisy datasets. To overcome this limitation, we propose a new multi-view graph learning framework, which for the first time simultaneously and explicitly models multi-view consistency and multi-view inconsistency in a unified objective function, through which the consistent and inconsistent parts of each single-view graph as well as the unified graph that fuses the consistent parts can be iteratively learned. Though optimizing the objective function is NP-hard, we design a highly efficient optimization algorithm which is able to obtain an approximate solution with linear time complexity in the number of edges in the unified graph. Furthermore, our multi-view graph learning approach can be applied to both similarity graphs and dissimilarity graphs, which lead to two graph fusion-based variants in our framework. Experiments on twelve multi-view datasets have demonstrated the robustness and efficiency of the proposed approach.

Automated machine learning (AutoML) is the sub-field of machine learning that aims at automating, to some extend, all stages of the design of a machine learning system. In the context of supervised learning, AutoML is concerned with feature extraction, pre processing, model design and post processing. Major contributions and achievements in AutoML have been taking place during the recent decade. We are therefore in perfect timing to look back and realize what we have learned. This chapter aims to summarize the main findings in the early years of AutoML. More specifically, in this chapter an introduction to AutoML for supervised learning is provided and an historical review of progress in this field is presented. Likewise, the main paradigms of AutoML are described and research opportunities are outlined.

In this paper, we propose a novel information criteria-based approach to select the dimensionality of the word2vec Skip-gram (SG). From the perspective of the probability theory, SG is considered as an implicit probability distribution estimation under the assumption that there exists a true contextual distribution among words. Therefore, we apply information criteria with the aim of selecting the best dimensionality so that the corresponding model can be as close as possible to the true distribution. We examine the following information criteria for the dimensionality selection problem: the Akaike Information Criterion, Bayesian Information Criterion, and Sequential Normalized Maximum Likelihood (SNML) criterion. SNML is the total codelength required for the sequential encoding of a data sequence on the basis of the minimum description length. The proposed approach is applied to both the original SG model and the SG Negative Sampling model to clarify the idea of using information criteria. Additionally, as the original SNML suffers from computational disadvantages, we introduce novel heuristics for its efficient computation. Moreover, we empirically demonstrate that SNML outperforms both BIC and AIC. In comparison with other evaluation methods for word embedding, the dimensionality selected by SNML is significantly closer to the optimal dimensionality obtained by word analogy or word similarity tasks.