Counterfactual Explanations for Time Series Data via Reinforcement Learning

ICLR 2026 Conference Submission20272 Authors

Keywords: Explainable AI (XAI), Counterfactual explanations, Reinforcement Learning, Time Series

Abstract: Counterfactual (CF) explanations are a powerful tool in Explainable AI (XAI), providing actionable insights into how model predictions could change under minimal input alterations. Generating CFs for time series, however, remains challenging: existing optimization-based methods are often instance-specific, impose restrictive constraints, and struggle to ensure both validity and plausibility. To address these limitations, we propose a reinforcement learning (RL) framework for counterfactual explanation in time series. Our actor–critic agent learns a policy in the latent space of a pre-trained autoencoder, enabling the generation of counterfactuals that balance validity and plausibility without relying on rigid handcrafted constraints. Once trained, the RL agent produces counterfactuals in a single forward pass, ensuring scalability to large datasets. Experiments on diverse benchmarks demonstrate that our approach generates valid and plausible counterfactuals, offering a reliable alternative to existing methods.

Primary Area: interpretability and explainable AI

Submission Number: 20272