A Self-explainable Model of Long Time Series by Extracting Informative Structured Causal Patterns

Explainability in AI systems needs to go beyond simple importance scores: Most current methods for explaining AI decisions about time series data only highlight which individual data points matter most, but fail to capture meaningful patterns like trends or cycles. This paper argues that better explanations should show how patterns unfold over time, not just which moments are important.

The framework balances accuracy with interpretability: EXCAP achieves strong predictive performance while simultaneously generating human-understandable explanations. This challenges the common assumption that more explainable AI models must sacrifice accuracy—suggesting responsible AI doesn't require this trade-off.

Causal reasoning strengthens trustworthiness in high-stakes applications: By grounding explanations in causal relationships (rather than just correlations), EXCAP is designed to be more reliable for critical domains like healthcare and finance, where understanding why a model made a decision is essential for human oversight and accountability.

Stable explanations over time build user confidence: The paper demonstrates that EXCAP produces explanations that remain consistent and robust even when the underlying data is slightly perturbed. This stability is important for trust—users are more likely to rely on AI systems whose reasoning doesn't dramatically shift with minor changes.

Structured pattern extraction makes AI reasoning more transparent: Rather than treating time series as black boxes, EXCAP explicitly identifies and explains coherent temporal segments and their causal relationships, making the model's decision-making process more accessible to non-technical stakeholders who need to validate or challenge AI recommendations.