Jundong Li, assistant professor of electrical and computer engineering at the University of Virginia, develops novel deep learning algorithms to glean more actionable patterns from graph-structured data sets. Li holds joint appointments in UVA’s Department of Computer Science and School of Data Science. Li and Shuiwang Ji, associate professor of computer science and engineering at Texas A&M, have earned a National Science Foundation grant to improve the essential building blocks of deep learning algorithms for graphs.
Graphs provide a general description of data and their relations. Graphs have been extensively used for modeling a plethora of real-world systems including social media platforms, collaboration networks, biological networks and critical infrastructure systems. Li and Ji’s research will propel state-of-the-art graph mining and deep learning research to a new frontier by taming complex real-world graphs, as well as advancing graph-related applications from different domains. Specifically, they focus on a cutting-edge technique called graph neural networks and aim to develop novel machine learning algorithms to improve the essential building blocks of graph neural networks by characterizing the properties of real-world graphs from different perspectives.
In this project, Li and Ji will first develop novel convolution operations to facilitate the node-level learning tasks on graphs. One specific problem they plan to tackle is the graph anomaly detection problem, which aims to find anomalous nodes whose patterns deviate significantly from other majority nodes. The problem has a broad spectrum of applications, ranging from social spam detection and network intrusion detection to system fault diagnosis. For example, by learning high-level representations of nodes, their developed convolution operations have the potential to identify the malicious users from a swarm of normal users in social media platforms. Later on, they also plan to develop novel pooling operations to support graph-level analytical tasks. For example, each chemical compound can be modeled as a graph and the developed pooling operations can be leveraged to predict the properties of different chemical compounds (e.g., toxic or not), which has significant implications in the areas of metabolic engineering and drug discovery.