Thursday, January 17, 2019
2:00 PM
Wegmans Hall 3201
Ph.D. Thesis Proposal
Shupeng Gui
University of Rochester
Universal and Robust Graph Representation Learning

Graph analysis is becoming a powerful machine learning tool for data mining, computer vision, natural language processing and other areas in the past few years. The blooming of applications of graph learning accelerates the development on graph analysis, especially Graph Representation Learning. Graph representation learning (GRL), also known as graph embedding, aims to transfer the structured data like graphs into the Euclidean space such as vectors. In fact, the solution of GRL is not trivial since there exist various types of graphs that they can be directed or undirected, static or dynamic, and homogeneous or heterogeneous. Meanwhile, the embedding problems are also probably under the malicious attack. Our goal is to propose a general embedding learning model for various types of graphs and carefully design a robust embedding learning mechanism against the adversaries.

First of all, we focus on the discussion of static graph embedding learning and propose a novel learning framework, P2IR. P2IR is a universal learning framework for both homogeneous and heterogeneous static graphs via the partial permutation invariant set functions. Theoretical guarantee for the representation capability of our method has been proved for general homogeneous and heterogeneous graphs. Evaluation results on benchmark data sets show that the proposed P2IR outperforms the state-of-the-art approaches on producing node vectors for classification tasks.

Beyond the static graphs, we are also interested in the dynamic graphs embedding learning. The proposed plan is to extend the learning framework of static graphs such as P2IR to the dynamic ones via applying the continuous time recurrent neural network (CTRNN) with the knowledge of the evolution of dynamic graphs. We also have a tentative plan for a robust graph embedding learning mechanism against the malicious graph data modification. We consider the adversarial problem under the white-box attack settings and the adversaries are able to slightly modify the connections, scales and attributes for a graph data set. As the literature review, such minor perturbations can lead to the performance corruption of deep neural networks. To avoid the effect of malicious attack behaviors, we involve the adversarial learning to setup a minimax optimization problem as the defensive learning mechanism.

Advisor: Prof. Ji Liu (Computer Science)

Committee: Prof. Daniel Gildea (Computer Science), Prof. Chenliang Xu (Computer Science), Prof. Gonzalo Mateos-Buckstein (Electrical and Computer Engineering)