Flow Matching Anomaly Detection Robust to Contaminated Training Data

This study investigates the application of flow matching techniques for unsupervised anomaly detection, specifically focusing on tabular data in financial settings where training datasets are often contaminated with anomalous samples. This contamination violates a core assumption of many anomaly detection methods, which typically expect clean, normal data. The research compares Time-Conditioned Contraction Matching (TCCM) with Forest-Flow, evaluating various anomaly scoring functions. A key finding is the critical importance of the chosen anomaly score. The original single-step Decision score used by TCCM proved highly sensitive to contaminated training data. In contrast, trajectory-based scores, such as Deviation and Reconstruction scores, provided significantly more stable anomaly signals. When these more robust scores are employed, Forest-Flow demonstrates competitive performance, in some cases even outperforming TCCM. These results underscore that for flow-matching methods in financial anomaly detection, especially with severe class imbalance, the anomaly scoring mechanism is paramount.