Go to AAAI 2025 Workshop AICT homepageTimeXAI: Unified Datasets and Concept-Based Counterfactual Explanations for Time Series
Keywords: Counterfactual Explanations, Explainability, Time series
TL;DR: A Concept-Based Counterfactual Explanations Framework for Time Series
Abstract: Explaining AI systems operating on time series data is crucial in many decision-making areas, such as healthcare, energy, and public policy-making, which requires interpretable and transparent explanations to overcome the black-box nature of models, especially for non-experts. Effective explanations allow us to understand how a model has learned, which helps in taking steps to improve robustness, safety, and fairness. Concept-based explanations have gained traction, offering insights into AI decisions using higher-level concepts. Concurrently, in our climate-conscious world, businesses increasingly rely on time series data to enhance energy efficiency and drive sustainable practices. Yet, several significant challenges persist. There is a lack of comprehensive archives for sustainable energy time series data, and current models often lack robust, regression-explainable methods to interpret their behavior. Our findings indicate that many existing models are prone to over-fitting specific open-source datasets, resulting in a disconnect between their performance in controlled environments and real-world applications. To address this, we introduced \method{TimeXAI}, a framework that uses counterfactual-based explanations to uncover these weaknesses and provide deeper insights into where and why models struggle. To further this effort, we introduce a comprehensive archive of 78 publicly available sustainable energy time series datasets, encompassing a total of over 137 million hourly instances at a 1Hz sampling rate. Our results strongly suggest that future work should explore more varied set time series to better assess model performance and prevent the risk of over-fitting to specific time series data sets. The archive and code can be accessed at \url{https://TimeXAI.github.io/}.
Submission Number: 20
Loading