Junxiang Wang

Junxiang Wang

Researcher

Selected Projects

Vaccine Adverse Event Detection on Social Media

A wide range of vaccinations have made a significant positive impact on global health. However, vaccinations can sometimes cause adverse reactions in a large population, which has become a significant healthcare issue. Severe reactions can even lead to death. To address the importance and potential consequences of adverse reactions, there is a need for a system that can quickly and accurately identify such events. The FDA’s Sentinel Initiative is an example of such a system. However, the traditional method of gathering Adverse Event(AE) reports relies on users submitting detailed descriptions through complex forms after recovering from the reactions. This system has two major drawbacks: a low number of formal reports due to complexity, and a time delay in submitting reports due to administrative processes. In contrast, social media platforms like Twitter and Facebook have become popular channels for sharing information, including healthcare-related updates.

Deep Learning Optimization by Alternating Minimization

Stochastic Gradient Descent (SGD) and its variants have gained popularity for training deep neural networks. However, theoretical guarantees on the convergence of SGD do not apply well to problems involving deep neural networks, which are nonsmooth and nonconvex. Additionally, SGD suffers from limitations such as the gradient vanishing problem, where the error signal diminishes during backpropagation, and sensitivity to poor conditioning, where small input changes lead to dramatic gradient changes. To address these drawbacks, Alternating Minimization(AM) methods have emerged as potential solutions for deep learning problems. These methods reformulate a neural network problem as a nested function with multiple linear and nonlinear transformations across layers. The nested structure is decomposed into a series of linear and nonlinear equality constraints using auxiliary variables and penalty hyperparameters. This decomposition generates subproblems that can be minimized alternately.

Graph Inverse Problems

Graphs are universal structures consisting of nodes and links. They have a wide range of applications in various domains such as social networks, biology, and recommender systems. Graph prediction problems involve predicting outcomes of nodes, links, or entire graphs. While these problems have been extensively studied, graph inverse problems, which involve backtracking graph inputs to achieve specific graph outputs, are less explored and represent open research areas. Graph inverse problems have significant applications, including graph source localization and network influence maximization. For instance, graph source localization identifies information sources in a graph that result in current information diffusion patterns, and it can be viewed as the reverse process of graph diffusion.

Nonconvex Generalization of Alternating Direction Method of Multipliers(ADMM)

Due to the advantages and popularity of non-differentiable regularized and distributive computing for complex optimization problems, the Alternating Direction Method of Multipliers (ADMM) has received a great deal of attention in recent years. While the ADMM has been extensively applied in the convex problems, its extension on the nonconvex problems are not well-studied. This motivates me to explore nonconvex problems addressed by ADMM.