Deep Representation Learning on Complex Graphs

Speaker: Cheng Zheng
Affiliation: Ph.D. Candidate

Via Zoom Only:    https://ucla.zoom.us/j/98377023978

Abstract:  Graph structural data is ubiquitous across research and application domains, encoding extensive information in social networks, protein interactional networks, citation networks, etc. Effective and efficient representation learning of graph structural information has played important roles in graph mining tasks like link prediction, node classification and clustering. However, there are great challenges from the complex graph structures, such as isomorphism, sparsity, and heterogeneity. To address the challenges, traditional approaches have employed user-defined heuristics like node degrees or kernel functions, which do not scale and fail to generalize to real large graphs.

In this dissertation, I will present the deep representation learning approaches that can generalize to different types of complex graphs. For intrinsic sparsity of sampled sequences from the input graph, we propose an adversarially regularized autoencoders (NetRA) to learn the node representations.  NetRA learns smoothly regularized vertex representations that well capture the network structure through jointly considering both locality-preserving and global reconstruction constraints.  For the real-life graphs with task-irrelevant edges, we present NeuralSparse and TSNet, which leverage supervised graph sparsification technique to improve generalization power by learning to remove potentially task-irrelevant edges from input graphs.  Our methods take both structural and non-structural information as input, utilizes deep neural networks to parameterize sparsification processes, and optimizes the parameters by feedback signals from downstream tasks.  We extensively evaluate our proposed models on multiple tasks with real-life complex graphs and the empirical results show the superior performance, generalization and scalability of proposed models.

Biography:  Cheng Zheng is a Ph.D. candidate in the UCLA Electrical and Computer Engineering Department under the supervision of Prof. Wei Wang and Prof. Jonathan Kao. He received the B.S. degree in physics from Tsinghua University in 2015 and the M.S. degree in computer science from UCLA in 2018. He held multiple internship positions at NEC Labs, Facebook Inc. and Google Inc. His research interests include graph mining, social network analysis and deep learning.

For more information, contact Profs. Jonathan Kao (kao@seas.ucla.edu) or Wei Wang (weiwang@cs.ucla.edu)

Date/Time:
Date(s) - Aug 04, 2020
2:00 pm - 4:00 pm

Location:
Map Unavailable