Achieving Counterfactual Explanation for Sequence Anomaly Detection

Abstract

We propose CFDet, a counterfactual explanation framework for sequence anomaly detection. CFDet identifies anomalous entries by generating minimal and plausible modifications that alter a model’s prediction from anomalous to normal. Experiments on BGL, Thunderbird, and CERT datasets demonstrate that CFDet produces high-fidelity explanations and consistently outperforms attention-based, Shapley value, and gradient-based baselines.