Evaluating explainability of time series models: A user-centered approach in industrial settings

Abstract

This paper investigates methods for evaluating the explainability of transformer models analyzing time series data, a largely unexplored area in the field of explainable AI (XAI). The study focuses on application-grounded methods involving human subject experiments with domain experts. On-site evaluations were conducted in two industrial settings involving 14 control room operators. The evaluation protocols consisted of methods to measure the metrics of subjective comparison, forward simulatability, and subjective satisfaction. The results indicate that the chosen combination of evaluation metrics provide a multi-faceted assessment on quality and relevance of explanations from an operator’s perspective in industrial settings, in turn contributing to the field of user-centered XAI evaluation, particularly in the context of time series data and offers insights for future work in this area.

Keywords: Explainable AI, Transformer Models, Time Series Data, Explainability Evaluation

DOI: 10.54941/ahfe1004651