Geometric Learning on Graph Structured Data

Abstract

Graphs provide a ubiquitous and universal data structure that can be applied in many domains such as social networks, biology, chemistry, physics, and computer science. In this thesis we focus on two fundamental paradigms in graph learning: representation learning and similarity learning over graph-structured data. Graph representation learning aims to learn embeddings for nodes by integrating topological and feature information of a graph. Graph similarity learning brings into play with similarity functions that allow to compute similarity between pairs of graphs in a vector space. We address several challenging issues in these two paradigms, designing powerful, yet efficient and theoretical guaranteed machine learning models that can leverage rich topological structural properties of real-world graphs.

This thesis is structured into two parts. In the first part of the thesis, we will present how to develop powerful Graph Neural Networks (GNNs) for graph representation learning from three different perspectives: (1) spatial GNNs, (2) spectral GNNs, and (3) diffusion GNNs. We will discuss the model architecture, representational power, and convergence properties of these GNN models. Specifically, we first study how to develop expressive, yet efficient and simple message-passing aggregation schemes that can go beyond the Weisfeiler-Leman test (1-WL). We propose a generalized message-passing framework by incorporating graph structural properties into an aggregation scheme. Then, we introduce a new local isomorphism hierarchy on neighborhood subgraphs. We further develop a novel neural model, namely GraphSNN, and theoretically prove that this model is more expressive than the 1-WL test. After that, we study how to build an effective and efficient graph convolution model with spectral graph filters. In this study, we propose a spectral GNN model, called DFNets, which incorporates a novel spectral graph filter, namely feedback-looped filters. As a result, this model can provide better localization on neighborhood while achieving fast convergence and linear memory requirements. Finally, we study how to capture the rich topological information of a graph using graph diffusion. We propose a novel GNN architecture with dynamic PageRank, based on a learnable transition matrix. We explore two variants of this GNN architecture: forward-euler solution and invariable feature solution, and theoretically prove that our forward-euler GNN architecture is guaranteed with the convergence to a stationary distribution.

In the second part of this thesis, we will introduce a new optimal transport distance metric on graphs in a regularized learning framework for graph kernels. This optimal transport distance metric can preserve both local and global structures between graphs during the transport, in addition to preserving features and their local variations. Furthermore, we propose two strongly convex regularization terms to theoretically guarantee the convergence and numerical stability in finding an optimal assignment between graphs. One regularization term is used to regularize a Wasserstein distance between graphs in the same ground space. This helps to preserve the local clustering structure on graphs by relaxing the optimal transport problem to be a cluster-to-cluster assignment between locally connected vertices. The other regularization term is used to regularize a Gromov-Wasserstein distance between graphs across different ground spaces based on degree-entropy KL divergence. This helps to improve the matching robustness of an optimal alignment to preserve the global connectivity structure of graphs. We have evaluated our optimal transport-based graph kernel using different benchmark tasks. The experimental results show that our models considerably outperform all the state-of-the-art methods in all benchmark tasks. Show more