EACE: Explain Anomaly via Counterfactual Explanations

Abstract

Anomaly detection aims to identify data points that deviate from the prevailing data distribution. Despite numerous anomaly detection models, there is a prevailing oversight in their interpretability, specifically regarding the rationale behind classifying a specific data point as an anomaly. Therefore, Interpretable Machine Learning has become a current research hotspot and is crucial for users to trust models. As one of the representative models, Counterfactual Explanation (CFE) methods generate alternative scenarios different from the observed data to explain model decisions. CFE tries to answer how the model’s output would change if certain factors (features) were altered. However, most existing CFE methods are designed for classification tasks, and it is a challenge for them to transform anomalies into counterfactual explanation samples effectively. To overcome this limitation, we propose a novel method for Explaining Anomaly via Counterfactual Explanation named EACE. Specifically, based on existing CFE methods’ limitations in handling anomalies, we propose a novel optimization objective by incorporating density loss and boundary loss. Meanwhile, we improved the genetic algorithm to solve this optimization problem since the new loss function is not differentiable. To evaluate the quality of the generated counterfactual explanations, we compare comprehensively with state-of-the-art counterfactual explanation methods and feature importance-based explanation methods. Experimental results demonstrate that EACE has a notable ability to convert anomalies into counterfactual explanation samples that are highly aligned with the normal cluster.