An Empirical Study of Explainable AI Techniques on Deep Learning Models For Time Series Tasks

XAI methods lack rigorous verification: Many explainable AI techniques currently used to interpret machine learning models produce outputs that haven't been properly verified or tested, relying instead on subjective human judgment of individual examples.

Time series explanations need specialized evaluation: Attribution methods developed for images and text don't automatically work well for time series data, highlighting the need for domain-specific approaches to AI explainability across different data types.

Automated testing can improve XAI reliability: The study proposes using perturbation methods—systematically removing or changing data to test explanations—as an objective way to automatically evaluate and rank which XAI techniques are actually trustworthy, rather than relying on visual inspection alone.

Benchmark frameworks are essential for responsible AI: Establishing standardized testing frameworks for XAI methods helps organizations identify which explanation techniques they can reliably depend on when deploying AI systems in real-world applications.

Explainability quality varies significantly: The research suggests that not all popular XAI techniques perform equally well, making it critical for practitioners to test and validate which methods work best for their specific use cases before using them to justify AI decisions.