EvtGraph Optimizes Temporal Graph Learning for Multimodal Time Series

Multimodal temporal data, such as clinical records or sensor streams, often exhibit irregular patterns and uneven information density. Traditional models that rely on uniform discretization struggle with this, leading to inefficient representations and computational overhead. EvtGraph proposes a novel framework to address these challenges by focusing on event-adaptive compression. The EvtGraph framework reparameterizes dense sequences into event-level tokens using an event-adaptive compression (EAMC) mechanism. It then selects a compact subset of these tokens based on a predefined node budget (NBC), effectively identifying the most salient events. Finally, it performs temporally constrained sparse graph reasoning (T2SG) on this reduced set of critical events. This design allows EvtGraph to allocate representational capacity efficiently under fixed budget constraints, transforming high-redundancy temporal data into structured computation over key events. Experimental results on clinical and cross-domain benchmarks show that EvtGraph significantly outperforms Transformer-based and recurrent baselines in both performance and efficiency, demonstrating that a small budget is often sufficient for practical applications.