October 1, 2026
Project overview
This project investigates whether groups of related card transactions reveal fraud missed by individual transaction scores, and whether their patterns can become short rules that detect future fraud. The available corpus contains 600 million labelled transactions, with 17 basic variables and approximately 150 variables after temporal enrichment.
A LightGBM baseline using all enriched variables provides the main comparison; a second baseline using only basic variables isolates the contribution of existing aggregates. Following a chronological evaluation protocol [1], students will examine transaction groups built from temporal proximity, shared entities and behavioural similarities [3]. Without independent campaign annotations, these groups remain candidate campaigns rather than confirmed fraud operations.
The next step is to extract interpretable rules from these groups and compare them with global rule ensembles [2]. Students will study whether rules generalise to later periods and previously unseen entities, while controlling false positives, redundancy and rule complexity. A further investigation asks whether a large language model supplied with variable definitions and contrasting group summaries can propose, simplify or combine better rules [7]. Proposals must follow an executable grammar and are evaluated against real labels; the language model operates offline.
Experiments will measure additional fraud recall beyond LightGBM under a common false-positive budget, together with rule stability, review effort and computational cost. Active learning and supervised score aggregation are optional extensions [4–6]. The expected outcome is a reproducible investigation explaining when groups and rules help or fail; a positive performance gain is not required.
Deliverables
- A reproducible prototype, candidate-group analysis and rule library with support and validity periods.
- A comparative report covering chronological evaluation, ablations, limitations and generation costs.
References
[1] Dal Pozzolo, A., Boracchi, G., Caelen, O., Alippi, C., & Bontempi, G. (2018). Credit Card Fraud Detection: A Realistic Modeling and a Novel Learning Strategy. IEEE Transactions on Neural Networks and Learning Systems, 29(8), 3784–3797. DOI: 10.1109/TNNLS.2017.2736643. Paper.
[2] Friedman, J. H., & Popescu, B. E. (2008). Predictive Learning via Rule Ensembles. The Annals of Applied Statistics, 2(3), 916–954. DOI: 10.1214/07-AOAS148. arXiv:0811.1679.
[3] Van Vlasselaer, V., Bravo, C., Caelen, O., Eliassi-Rad, T., Akoglu, L., Snoeck, M., & Baesens, B. (2015). APATE: A Novel Approach for Automated Credit Card Transaction Fraud Detection Using Network-Based Extensions. Decision Support Systems, 75, 38–48. DOI: 10.1016/j.dss.2015.04.013. Paper.
[7] Zhang, H., & Jain, S. (2026). LLM-Assisted Logic Rule Learning: Scaling Human Expertise for Time Series Anomaly Detection. Workshop Time Series in the Age of Large Models, ICLR 2026. arXiv:2601.19255.
[8] Lusis Chair – LISN – CentraleSupélec. Chair website and fraud detection research context.
Further reading for optional extensions
[4] Das, S., Wong, W.-K., Dietterich, T. G., Fern, A., & Emmott, A. (2016). Incorporating Expert Feedback into Active Anomaly Discovery. IEEE ICDM, 853–858. DOI: 10.1109/ICDM.2016.0102. Paper.
[5] Carcillo, F., et al. (2018). Streaming Active Learning Strategies for Real-Life Credit Card Fraud Detection: Assessment and Visualization. International Journal of Data Science and Analytics. DOI: 10.1007/s41060-018-0116-z. arXiv:1804.07481.
[6] Bleakley, K., et al. (2026). Supervised Score Aggregation for Active Anomaly Detection. Transactions on Machine Learning Research, January 2026. Paper. Code.