Inductive Graph Inference: Challenges and Solutions

The Hong Kong University of Science and Technology
Department of Computer Science and Engineering


PhD Thesis Defence


Title: "Inductive Graph Inference: Challenges and Solutions"

By

Mr. Meng QIN


Abstract:

Most graph inference tasks have transductive and inductive settings, where 
inductive inference is a more advanced setting that can transfer knowledge 
learned from existing known topology to new nodes or graphs. However, the 
original designs of many graph inference techniques only consider 
transductive settings. Some of them are also inapplicable to their inductive 
settings. This thesis studies several challenging inductive inference tasks 
seldom considered in related research, covering the inference for new nodes 
and across graphs as well as extended inductive inference without graph 
neural networks (GNNs).

First, we consider inductive temporal link prediction for weighted graphs 
with non-fixed node sets and develop an inductive dynamic embedding 
aggregation (IDEA) method based on designs of a stacked GNN-RNN cell, an 
adaptive embedding aggregation module, and a hybrid training objective 
regarding adversarial learning and scale difference minimization. Experiments 
verify that IDEA can handle the node variation while deriving high-quality 
prediction results for weighted graphs.

Second, we consider inductive community detection across graphs. Two methods 
(i.e., ICD and PRoCD) are proposed to handle graphs from the same and 
different domains, respectively. For both methods, we first conduct the 
offline (pre-)training of a GNN on known graphs and generalize it to new 
graphs for fast online inference without re- training. A better trade-off 
between quality and efficiency can be achieved during the online inference on 
new graphs.

Third, we consider the joint inductive inference of identity and position 
embeddings without using the attribute aggregation of GNNs. An inductive 
random walk embedding (IRWE) method is proposed. It combines multiple 
attention units to handle the random walk on graph topology and 
simultaneously derives the two types of embeddings. We demonstrate that there 
exist topology-based features (e.g., anonymous walk and induced statistics) 
shared by all possible graph topology that can be used to support inductive 
inference.


Date:                   Monday, 2 December 2024

Time:                   10:00am - 12:00noon

Venue:                  Room 3494
                        Lifts 25/26

Chairman:               Dr. Yang LU (ECON)

Committee Members:      Prof. Dit-Yan YEUNG (Supervisor)
                        Dr. Yangqiu SONG
                        Prof. Raymond WONG
                        Prof. Wing Hung KI (ECE)
                        Prof. Irwin Kuo Chin KING (CUHK)