Temporal Relational Graph Learning

Abstract

This paper introduces a novel learning algorithm, Temporal Relational Graph Learning (TRGL), designed to effectively capture and leverage relational structure within dynamic graphs, explicitly incorporating temporal dependencies. TRGL addresses the limitations of existing approaches by employing a "relational memory" module – a recurrent neural network integrated with a differentiable relational tensor – to dynamically update and refine relationship representations based on observed temporal changes. The core mechanism utilizes a recurrent network to process the graph's evolving state, while the relational tensor allows for the representation and modification of relationships. We demonstrate that this approach moves beyond static graph embeddings and traditional temporal graph networks by directly modeling the evolution of relationships and their interdependencies, resulting in a more robust representation of dynamic systems. The algorithm's performance is evaluated through theoretical analysis and conceptual design, outlining its potential for applications in areas such as social network analysis, anomaly detection, and dynamic systems modeling.