Interpretable Contrastive Learning for Robust and Explainable Anomaly Detection in Financial and Organizational Data

Anomaly detection is a fundamental task in modern intelligent systems, yet achieving both robustness and interpretability remains a persistent challenge, especially when labeled data are scarce. Recent advances in contrastive learning have enabled effective representation learning from unlabeled data, but most existing models offer limited insight into the feature-level factors underlying their predictions. In this paper, we propose an interpretable contrastive learning framework for robust and explainable anomaly detection. By integrating a learnable feature attribution mask at the input level directly into the contrastive objective, our approach not only enhances detection accuracy but also provides faithful, sparse attributions for each detected anomaly. Unlike prior interpretable anomaly detection methods, our framework jointly optimizes representation learning and attribution faithfulness, offering both theoretical generalization guarantees and practical scalability for real-world deployment. We conduct comprehensive experiments on four benchmark datasets spanning network intrusion detection, organizational communication analysis, financial fraud identification, and synthetic banking scenarios, and compare our method to both classical and state-of-the-art baselines. The results demonstrate consistent improvements, with our approach achieving up to 4–7% higher AUC than recent deep learning baselines, along with high stability of explanations and clear visual separation of anomalies in the learned embedding space. Beyond these quantitative gains, our work underscores the importance of actionable interpretability, providing a principled and deployable solution for anomaly detection in high-stakes domains such as finance, IoT, and cyber-security.