A Deep Dive into Perturbations as Evaluation Technique for Time Series XAI

Evaluating AI explanations is harder than creating them: The paper addresses a critical gap in explainable AI—while many techniques exist to explain why AI models make decisions, there's no standard way to verify whether those explanations are actually good or trustworthy, particularly for time series data used in finance, healthcare, and climate science.

Perturbation testing can validate explanations: The researchers demonstrate that systematically changing input data and observing how explanations respond is an effective method to assess explanation quality. This approach helps determine which XAI techniques reliably identify the most important factors influencing model decisions.

Different XAI methods have different strengths: The study found that various explanation techniques perform differently depending on what you're measuring—some are better at identifying which data points matter most ("precision"), while others are better at quickly detecting when changes occur ("return time"), suggesting no single best approach exists for all situations.

Practical guidance for selecting explanation methods: The findings provide concrete recommendations for practitioners choosing XAI techniques for time series analysis, emphasizing that selection should depend on specific use-case priorities rather than assuming all methods are equally reliable.

Better explanations support responsible AI deployment: By establishing rigorous evaluation methods for AI explanations, this work helps ensure that when AI systems are used in high-stakes domains like healthcare and finance, stakeholders can actually trust and verify the reasoning behind critical decisions.