Learning Approach for Querying Neural Graph Databases

PhD Thesis Proposal Defence


Title: "Learning Approach for Querying Neural Graph Databases"

by

Mr. Zihao WANG


Abstract:

Neural Graph Databases (NGDB) augment classic graph databases with neural 
representations that embed both features and relations, but this integration 
also introduces new and complex challenges in handling logical queries. The 
logical calculus required to answer queries is not readily found in continuous 
spaces, making querying such NGDB a highly intricate and non-trivial task. An 
important example of neural graph databases is knowledge graph with relation 
and entity representations, which are then our key research object.

The first part of this thesis discusses learning approaches for tree-formed 
queries. For such queries, the logical queries are solved by set operators over 
geometric embedding spaces of a fixed dimension, where only constant data 
complexity is required. We introduce the large-scale query-answering framework, 
demonstrating the challenge of generalizing such approaches in the 
combinatorial spaces of logical queries due to the conflict between rigorous 
logical reasoning and geometric embeddings. To overcome this, we first 
understand the truth value matrix between queries and all entities from a 
matrix-decomposition perspective. We then show that the fuzzy logical t-norm 
operations can be integrated into the embeddings under specific conditions. 
This leads to developing a new metric space inspired by unbalanced optimal 
transport for more combinatorically generalizable query answering and its 
computation with convolution operations.

In the second part of the thesis, we take one step back from the Tree Form (TF) 
queries and turn to Existential First Order (EFO) queries. We present 
syntactical and complexity gaps between the TF and EFO queries, which motivate 
further research in learning methods beyond tree-formed queries. With an 
abstract interface of link predictors, the query answering process of EFO query 
can be regarded as an optimization process but requires at least linear data 
complexity. We then propose a learning-to-optimization framework to answer such 
queries. Compared to direct global optimization, we introduce four-types the 
one-hop inference of single predicate whose results are then combined by a 
graph neural network for global solution. Notably, we derive the closed-form 
formulations of the one-hop inferences for widely used six KG embeddings, which 
further facilitate the inference process.


Date:                   Friday, 2 August 2024

Time:                   10:00am - 12:00noon

Venue:                  Room 5506
                        Lifts 25/26

Committee Members:      Dr. Yangqiu Song (Supervisor)
                        Prof. Dit-Yan Yeung (Chairperson)
                        Prof. Raymond Wong
                        Prof. Xiaofang Zhou