Learning with time-varying heterogeneous graphs
Computer scientist at Information Sciences Institute,
University of Southern California
COM2 Level 1
SR@LT19
Abstract:
Many big data are naturally mapped to heterogeneous graphs (networks) with different types of nodes (e.g., users, images, documents) and relations (e.g., friendship, ownership, clicking, and viewing). In order to turn big data to value, we first need to understand graphs representing data. But, how can we learn from graphs through analysis? Are the graph analytical algorithms working well in the real-life streaming scenario (e.g., a continuous stream of networks with millions of new links between different types of objects is received per hour)? In this talk, I will present how our heterogeneous graph modeling and scalable graph learning algorithms help answer these questions and make sense of large-scale time-varying heterogeneous graphs.
Concretely, I will focus on two of my works: tripartite graph clustering for sentiment analysis and multi-type graph summarization for entity resolution. Using these two works as examples, I will illustrate how different data analytic tasks can be formulated as graph problems. For example, I will explain the novel tripartite model that represents twitter data with different sources and transforms the sentiment analysis task into a graph-clustering problem. To meet the timeliness requirement of processing large streaming graph data, I will also describe scalable algorithms that incrementally learns the tripartite clustering. Finally, I will also briefly describe some of my other works on large-scale network analysis.
Biodata:
Linhong Zhu is currently a Computer scientist at Information Sciences Institute, University of Southern California and a senior software engineer at Pinterest. She obtained her Ph.D. degree in computer engineering from Nanyang Technological University, Singapore in 2011. Her research interests are large-scale graph analytics with applications to social network analysis, semantic web, and predictive modeling. Her papers are published in prestigious international journals (such as ACM TODS, IEEE TKDE) and proceedings (such as ACM SIGMOD, ACM SIGKDD, ACL, ICDE, ICDM and ISWC). Her work has been awarded as the best research paper award in ISWC 2016 and has been selected as two of the best papers in SIGMOD 2010. She has been awarded with University of Southern California Postdoctoral travel and training award in 2014 and served as Co-PI for several DARPA projects.