TSFeatLIME: An Online User Study in Enhancing Explainability in Univariate Time Series Forecasting

Explainability benefits non-technical users most: The user study of 160 participants found that explanations generated by TSFeatLIME were "significantly more effective for participants without a computer science background," suggesting that improved AI explanations can democratize understanding of complex forecasting models across diverse audiences.

Better explanations require considering data relationships: TSFeatLIME improves upon existing methods by incorporating "pairwise Euclidean distances between the queried time series and the generated samples," demonstrating that explanations become more trustworthy when they account for how similar new data is to training examples.

Surrogate models can make black-box predictions more understandable: The framework successfully enables surrogate models to "better simulate the behaviour of the black-box" while maintaining accuracy, offering a practical approach to making opaque forecasting systems more transparent without compromising their performance.

User testing is essential for validating explainability claims: Rather than assuming explanations are helpful, the researchers conducted an interactive user study measuring whether participants could actually predict model behavior changes, highlighting that responsible AI requires empirical validation that explanations genuinely improve human understanding.