DIDER: Discovering Interpretable Dynamically Evolving Relations

Effective understanding of dynamically evolving multiagent interactions is crucial to capturing the underlying behavior agents in social systems. It usually challenging observe these directly, and therefore modeling latent essential for realizing complex behaviors. Recent work on Dynamic Neural Relational Inference (DNRI) captures explicit inter-agent at every step. However, prediction step results noisy lacks intrinsic interpretability without post-hoc inspection. Moreover, it requires access ground truth annotations analyze predicted interactions, which are hard obtain. This letter introduces DIDER, Discovering Interpretable Dynamically Evolving Relations, a generic end-to-end interaction framework with interpretability. DIDER discovers an interpretable sequence by disentangling task into sub-interaction duration estimation. By imposing consistency type over extended time duration, proposed achieves requiring any We evaluate both synthetic real-world datasets. The experimental demonstrate that disentangled dynamic relations improves performance trajectory forecasting tasks.