Delta-XAI: A Unified Framework for Explaining Prediction Changes in Online Time Series Monitoring

Existing AI explanation methods often miss the full picture in time series monitoring: Most current explainability techniques analyze each moment in time independently, failing to account for how past events influence predictions—a critical oversight in sensitive domains like healthcare and finance where temporal patterns matter.

Explaining why predictions change is harder than explaining predictions themselves: The paper identifies that current XAI methods struggle specifically with explaining prediction changes over time, which is essential for practitioners to understand when and why a model's decisions shift, yet this challenge has been largely overlooked.

Simpler, older methods can outperform newer approaches when properly adapted: Classical gradient-based explanation techniques like Integrated Gradients performed better than recent specialized methods when adapted for temporal analysis, suggesting that the right framework matters more than algorithmic novelty.

A new method (SWING) better captures how past data influences current predictions: The proposed Shifted Window Integrated Gradients approach systematically incorporates historical observations to explain predictions more faithfully, reducing misleading explanations caused by unrealistic data scenarios.

Rigorous evaluation frameworks are essential for trustworthy AI explanations: The paper introduces a comprehensive evaluation suite assessing multiple dimensions (faithfulness, sufficiency, coherence) rather than single metrics, highlighting that responsible AI requires multi-faceted validation of explanation quality before deployment in high-stakes domains.