TIMING: Temporality-Aware Integrated Gradients for Time Series Explanation

Current explanation methods miss half the story: Conventional explainability tools for time series AI focus only on the magnitude of feature importance but ignore whether those features push predictions up or down. This means they can miss critical moments where data points actively work against a model's prediction—an important distinction for understanding AI decision-making.

Existing evaluation approaches hide problems: The paper reveals that standard metrics for assessing explanation quality inadvertently cancel out positive and negative contributions, making flawed explanation methods appear more reliable than they actually are. The authors propose new evaluation metrics (CPD and CPP) that separately measure whether methods correctly identify points with opposing directional impacts.

Temporal context matters for fair explanations: Simply applying existing explanation techniques to time series data can produce misleading results because these methods generate artificial data paths that ignore how time series naturally evolve and can create unrealistic scenarios. TIMING addresses this by maintaining temporal relationships, ensuring explanations reflect how the AI actually processes sequential information.

Better explanations are achievable without sacrificing theory: Rather than abandoning established explanation methods, TIMING improves upon them by adding temporal awareness while preserving their mathematical foundations, suggesting that responsible AI improvements don't require starting from scratch but can build thoughtfully on existing approaches.

Rigorous testing confirms practical improvements: Experiments on both synthetic and real-world datasets demonstrate that TIMING outperforms existing time series explanation methods, providing evidence that these theoretical improvements translate to more reliable explanations in practice.