abstract
- Explainability in outlier detection is a crucial requirement in high-risk domains such as medicine and finance. As models increase complexity to improve accuracy, their interpretability is often hindered, creating a significant trade-off. Furthermore, the ability to handle both numerical and categorical attributes within the same model remains a challenge. To address this, we propose the Explainable Outlier Tree-based Encoder (EOTE), a novel anomaly detection model that integrates classification and regression trees within an autoencoder framework. EOTE generates human-readable explanations of outlier scores and can learn from mixed-attribute datasets. We evaluate EOTE against 12 leading anomaly detection algorithms across 110 datasets with mixed or single attribute type data. Our findings show that EOTE is one of the top-performing algorithms at detecting outliers in datasets with a single data type (numerical and nominal) and with mixed attribute data. Additionally, our proposal is able to provide explanations, making it suitable for use in high-risk applications. © 2025 Elsevier Ltd