Exploring Local Information for Graph Representation Learning

Graphs are important data structures that can capture interactions between individual entities. The primitive graph representation is usually high-dimensional, sparse, noisy, and in irregular forms, which is challenging for direct usage in downstream tasks (e.g., node or graph classification). Various graph representation learning (GRL) techniques have been developed to convert the raw graph data into low-dimensional vector representations while preserving the intrinsic graph properties. Graph neural networks (GNNs) are currently the most popular paradigm that can utilize nodes’ local information to assist their representation learning. Local neighborhood information in graphs varies greatly for different nodes. Therefore, direct neighborhood aggregation in GNNs is not an optimal choice. This thesis aims to develop new GNNs by exploring and modeling different types of local information to tackle the weaknesses of current GNNs in both algorithms and applications. We first propose three new models: 1) node feature convolution for graph convolutional network (NFC-GCN) to consider feature-level attention of local information; 2) learnable aggregator for GCN (LA-GCN) to generalize NFC-GCN further by lifting constraints on the input data format; and 3) hop-hop relation-aware GNN (HHR-GNN) to incorporate hop-level attention of local information. Moreover, we apply HHR-GNN to two industrial graphs for personalized video search and cross-domain recommendation tasks. Experimental studies show that the proposed methods have outperformed related state-of-the-art methods in both standard tasks of node/graph classification as well as application-specific tasks of search and recommendation.