Enhancing Graph Representation Learning: Data-Aware Advances in Graph Neural Networks

Abstract

Graph-structured data is a fundamental data modality across diverse domains, in which the relations among entities can be as crucial as the entities themselves. However, the irregular nature of it renders traditional Euclidean-based deep learning methods inadequate for modeling its intricate interdependent structure. Graph Neural Networks (GNNs) have emerged as the de facto paradigm for learning effective representations that capture relational information, offering favorable performance through their message-passing mechanism. Despite their development, many recent advances address different tasks primarily in a model-centric view, revolving around more sophisticated designs to enhance representation learning. Considering that the evolving demand for classification further drives models toward more complex architectures, this thesis aims to contribute to the advancement of effective representation learning by placing data at the center of methodological design. To this end, the thesis delves into various data-aware schemes to efficiently and delicately exploit the potential of the graph, thereby enhancing the graph representation learning capabilities of GNNs.