Why Do Class-Dependent Evaluation Effects Occur with Time Series Feature Attributions? A Synthetic Data Investigation

Perturbation-based evaluation metrics—commonly used to assess AI explanation quality when ground truth is unavailable—can produce unreliable and contradictory results depending on which class is being predicted. The paper demonstrates that these "class-dependent evaluation effects" occur even in simple, controlled scenarios, raising concerns about whether current industry-standard evaluation practices actually measure what they claim to measure.

Ground truth and perturbation-based evaluation methods frequently disagree on which attribution methods work best, with only weak correlations between the two approaches. This means organizations relying solely on perturbation metrics to validate their explainability tools may be making decisions based on misleading assessments, potentially deploying unreliable explanations to stakeholders.

Class-dependent effects emerge from basic differences between classes—such as variations in feature strength or timing—rather than from complex model behavior. This suggests the problem is fundamental to how these evaluation methods work, not a rare edge case, and may affect many real-world applications across different domains.

Current XAI evaluation practices may not be trustworthy enough to reliably guide method development and selection. The paper's findings suggest researchers and practitioners should be cautious about using perturbation-based metrics as the sole basis for choosing which explanation methods to deploy in high-stakes applications.