Enhancing Explainable Anomaly Detection: A Multi-Dimensional Entropy-based Approach

As automated anomaly detection becomes increasingly prevalent in large-scale data systems, understanding the causes of anomalies is critical for effective decision-making. Traditional anomaly detection models often operate as black boxes, providing little insight into the factors driving abnormal behavior. This study enhances Exstream, an entropy-based anomaly explanation framework, by improving feature selection and generalizing entropy computations for multi-dimensional feature interactions. We explore the potential of latent-space sampling to refine normal point selection and investigate alternative clustering methods to improve anomaly explanation robustness. Our findings demonstrate that while domain dependency limits latent-space sampling effectiveness, generalizing entropy across multiple features enhances the interpretability of anomalies in complex datasets. The proposed modifications offer a step forward in creating more transparent and actionable explanations for anomaly detection in time-series data. Our proposed model achieves 0.89 precision, 0.85 recall, an F1-score of 0.87, and an AUC-ROC of 0.91 on benchmark datasets (KDD 1999, UNSW-NB15), clearly demonstrating the model’s effectiveness and explainability compared to existing methods.