October 1, 2025
Project overview
This project, conducted within the Lusis Chair on AI and Finance, investigates anomaly detection in credit card payment data. Fraudulent transactions represent less than 0.5% of the available examples, making classification highly imbalanced. Payment records also combine continuous variables, such as transaction amounts, with categorical variables, such as merchant codes, whose representation can substantially increase dimensionality.
Gradient boosted trees (GBT) provide a strong baseline for tabular data, both in predictive performance and computational efficiency. The project explores whether a recent deep learning approach based on flow matching can offer a useful alternative, balancing detection quality against training time and the inference speed required for payment processing.
The starting point is Time-Conditioned Contraction Matching (TCCM) [1]. This method learns a time-conditioned vector field that contracts normal samples toward a fixed target. Transactions receive an anomaly score from their deviation from the expected contraction, using a single forward pass rather than numerical integration of a complete trajectory. Its formulation also supports feature-level attribution of anomaly scores, an interesting property for understanding suspicious payments.
Students will implement and analyse the method, first on standard public datasets and then on a private dataset of labelled bank transactions provided for the project. Experiments will compare TCCM with GBT, examining classification quality alongside the computational complexity of training and inference. The study should identify the method’s limitations and propose improvements suited to fraud detection.
Deliverables
- A Python/PyTorch implementation and experiments on public and private datasets.
- A comparative report covering detection performance, training cost, inference speed, limitations and proposed improvements.
References
[1] Li, Z., Huang, Q., Zhu, Y., Yang, L., Amiri, M. M., van Stein, N., & van Leeuwen, M. (2025). Scalable, Explainable and Provably Robust Anomaly Detection with One-Step Flow Matching. Advances in Neural Information Processing Systems, 38 (NeurIPS 2025). https://doi.org/10.52202/085713-2936. arXiv:2510.18328.