Consistent Counterfactual Explanations via Anomaly Control and Data Coherence

—Algorithmic recourses are popular methods to provide individuals impacted by machine learning models with recommendations on feasible actions for a more favourable prediction. Most of the previous algorithmic recourse methods work under the assumption that the predictive model does not change over time. However, in reality, models in deployment may both be periodically retrained and have their architecture changed. Therefore, it is desirable that the recourse should remain valid when such a model update occurs, unless new evidence arises. We call this feature consistency. This paper presents Anomaly Control and Data Coherence (ACDC), a novel model-agnostic recourse method that generates counterfactual explanations, i.e., instance-level recourses. ACDC is inspired by anomaly detection methods and uses a one-class classifier to aid the search for valid, consistent, and feasible counterfactual explanations. The one-class classifier asserts that the generated counterfactual explanations lie on the data manifold and are not outliers of the target class. We compare ACDC against several state-of-the-art recourse methods across four datasets. Our experiments show that ACDC outperforms baselines both in generating consistent counterfactual explanations, and in generating feasible and plausible counterfactual explanations, while still having proximity measures similar to the baseline methods targeting the data manifold. Code is available at https://github.com/mmovin/acdc. Impact Statement—Our paper addresses the challenge of generating consistent counterfactual explanations (CF) in the context of machine learning models. While existing algorithmic recourse methods assume a static model, our novel approach, Anomaly Control and Data Coherence (ACDC), maintains validity even when the model is retrained or its architecture changes. By leveraging a one-class classifier and adherence to the data manifold, ACDC produces CF explanations that are both valid and coherent. We conducted experiments comparing ACDC with state-of-the-art methods, demonstrating superior performance in generating consistent and feasible CF explanations. Furthermore, we explored the benefits and drawbacks of incorporating model information into the method and proposed a technique for training ACDC using sampled data in the absence of labeled training data. Our findings contribute to trustworthy decision systems and offer practical solutions for individuals affected by machine learning predictions, ensuring the persistence of valid and actionable recourse in dynamic model environments.