Temporal Graph Neural Networks with Adaptive Relational Strength

Abstract

This paper introduces a novel approach to Graph Neural Networks (GNNs) designed to effectively capture temporal dependencies within dynamic graph structures. Traditional GNNs often fail to adequately represent the evolving nature of relationships between nodes, hindering their performance in systems where the graph topology changes over time. Our proposed Temporal Graph Neural Networks (T-GNNs) address this limitation by incorporating a 'temporal strength' factor into node embeddings. This factor is dynamically adjusted based on the observed evolution of relationships within the graph, learned through a recurrent attention mechanism that specifically focuses on relational changes. The core innovation lies in the adaptive adjustment of relational strength, allowing the network to prioritize relevant temporal information. We demonstrate through a theoretical analysis and structural design that this approach significantly improves the ability of GNNs to model time-evolving graph data. The resulting T-GNN architecture offers a robust framework for analyzing and predicting behavior in dynamic systems represented as graphs.