Temporal Self-Explanation: Interpretable and Efficient Representation Learning for Dynamic Graphs

While models like DySAT (Sankar et al., 2020) and GDGNN (Kong et al., 2022) have pushed the envelope for dynamic and efficient graph representation learning, they often do so at the expense of transparency—it's unclear how or why certain temporal patterns are captured. This research proposes a new class of temporal graph neural networks equipped with built-in self-explanation modules (e.g., attention weight visualizations, temporal attribution maps, or counterfactual path tracing). These models would not only learn useful node/graph representations over time but also provide human-interpretable explanations for their outputs. The novelty lies in integrating interpretability into the core learning paradigm, rather than as a post-hoc add-on. This approach builds on recent trends toward explainable AI and directly addresses the need for trust and accountability in dynamic, real-world applications (like social networks, recommendation systems, or protein folding). The impact would be both practical (more actionable insights) and scientific (better understanding of how temporal dependencies are represented).

References:

1. Geodesic Graph Neural Network for Efficient Graph Representation Learning. Lecheng Kong, Yixin Chen, Muhan Zhang (2022). Neural Information Processing Systems.
2. DySAT: Deep Neural Representation Learning on Dynamic Graphs via Self-Attention Networks. Aravind Sankar, Yanhong Wu, Liang Gou, Wei Zhang, Hao Yang (2020). Web Search and Data Mining.