Introducing the Attribution Stability Indicator: a Measure for Time Series XAI Attributions

Explainability tools for time series models need reliability checks: The paper addresses a critical gap in responsible AI by proposing the Attribution Stability Indicator (ASI), a method to evaluate whether explanations from time series models are trustworthy and robust—an important consideration for high-stakes domains like healthcare, finance, and weather forecasting where decisions based on model explanations could have significant consequences.

Current explanation methods lack standardized quality measures: The research highlights that while attribution techniques can extract explanations from complex time series models, there is no established way to assess whether these explanations are actually reliable. This means organizations using these tools may be making decisions based on explanations that haven't been properly validated.

The ASI method tests explanation stability through perturbation analysis: The proposed measure works by slightly modifying input data and checking whether the model's explanations remain consistent and meaningful—similar to asking "if I change the data slightly, do the explanations still make sense?" This approach helps identify whether explanations reflect genuine patterns or are artifacts of the model.

Robustness and trustworthiness are measurable properties of AI explanations: Rather than treating explainability as a binary feature (either present or absent), this work demonstrates that explanation quality can be quantified and compared across different attribution techniques, enabling organizations to select more reliable explanation methods for their applications.