Inductive Graph Inference: Challenges and Solutions

PhD Thesis Proposal Defence


Title: "Inductive Graph Inference: Challenges and Solutions"

by

Mr. Meng QIN


Abstract:

For various complex systems, graphs serve as a generic abstract of the 
available entities and associated relations using a set of nodes and edges. 
Most graph inference tasks have both transductive and inductive settings. 
Compared with transductive inference, which only focuses on currently observed 
graph information, inductive graph inference is a more advanced setting that 
can transfer knowledge learned from known topology to new unseen nodes or 
graphs. However, the original designs of several graph inference techniques 
(e.g., link prediction and community detection) only focus on transductive 
settings. Some of them are also inapplicable to their inductive settings. This 
thesis studies several challenging inductive graph inference tasks seldom 
considered in related research, which covers both inference settings for new 
unseen nodes and across graphs.

First, we consider inductive temporal link prediction (TLP) for weighted 
dynamic graphs with non- fixed node sets, which should (i) determine the 
existence of future links, (ii) predict associated link weights, and (iii) deal 
with the variation of node sets. To support this challenging task, we develop 
an inductive dynamic embedding aggregation (IDEA) method based on novel designs 
of (i) a stacked GNN-RNN cell, (ii) an adaptive embedding aggregation module, 
and (iii) a hybrid training objective regarding adversarial learning and scale 
difference minimization. Experiments on public datasets of various scenarios 
demonstrate that IDEA can handle the variation of node sets and derive 
high-quality prediction results for weighted dynamic graphs.

Second, we consider community detection (CD) and explore the potential of an 
inductive scheme across graphs to achieve a better trade-off between the 
inference quality and efficiency of CD. Following this scheme, we propose an 
inductive community detection (ICD) method. It first (i) conducts the offline 
training of an advanced graph neural network (GNN) model on existing known 
graphs and then (ii) generalizes the model to new graphs for fast online 
inference without retraining. Novel designs of (i) an adversarial dual GNN and 
(ii) a clustering regularization objective also enable ICD to capture the 
permutation-invariant label information of CD during the offline training. 
Experiments on various synthetic and real benchmarks demonstrate that ICD can 
achieve a significant trade-off between the inference quality and efficiency of 
CD.


Date:                   Monday, 27 May 2024

Time:                   3:00pm - 5:00pm

Venue:                  Room 4475
                        Lifts 25/26

Committee Members:      Prof. Dit-Yan Yeung (Supervisor)
                        Prof. Raymond Wong (Chairperson)
                        Dr. Long Chen
                        Dr. Dan Xu