CLOUD-NATIVE DISTRIBUTED SYSTEMS FOR REAL-TIME PAYMENT INTELLIGENCE
Keywords:
Real-time payment intelligence, Fraud detection, Cloud-native systems, Microservices, Online learning, latency, Class imbalance, ReproducibilityAbstract
Cloud-native payment platforms increasingly route every transaction through an online fraud-scoring microservice whose response time directly affects the end-to-end approval latency seen by merchants and cardholders. In this work we present a controlled empirical study of seven CPU-friendly classifiers for real-time payment fraud detection, evaluated on the public credit-card transaction dataset released by the ULB Machine Learning Group and Worldline. Models are trained on the first 60% of the transaction stream in chronological order and evaluated on the subsequent 40%, with thresholds tuned on an intermediate validation window under a precision-floor service-level constraint. For each detector we jointly report ranking quality (ROC-AUC, PR-AUC), operating-point precision and recall, single-row inference latency at the median, 95th and 99th percentiles, batch throughput, and serialized model footprint. We find that a regularized logistic-regression baseline attains a PR-AUC of 0.74 with a sub-millisecond p99 single-row latency and a 2 KB on-disk footprint, while a 50-tree random forest achieves the highest PR-AUC (0.81) at roughly an order of magnitude higher p99 latency and several hundred times the footprint. We further examine a sample-weighted incremental update of an online stochastic-gradient model over the validation stream and observe that, on this benchmark, simple incremental updates can shift the decision boundary in a way that increases false-positive load without improving recall, motivating more careful update strategies in production cloud-native pipelines. The implementation package can be released to support reproducibility.References
[1] Jamshidi P, Pahl C, Mendonça N C, et al. Microservices: The journey so far and challenges ahead. IEEE Software, 2018(3): 24–35.
[2] Dean J, Barroso L A. The tail at scale. Communications of the ACM, 2013(2): 74–80.
[3] Dal Pozzolo A, Caelen O, Johnson R A, et al. Calibrating probability with undersampling for unbalanced classification. Proc. IEEE Symposium Series on Computational Intelligence (SSCI), 2015: 159–166.
[4] Dal Pozzolo A, Boracchi G, Caelen O, et al. Credit card fraud detection: A realistic modeling and a novel learning strategy. IEEE Transactions on Neural Networks and Learning Systems, 2018(8): 3784–3797.
[5] Dal Pozzolo A, Caelen O, Le Borgne Y A, et al. Learned lessons in credit card fraud detection from a practitioner perspective. Expert Systems with Applications, 2014(10): 4915–4928.
[6] Carcillo F, Dal Pozzolo A, Le Borgne Y A, et al. SCARFF: A scalable framework for streaming credit card fraud detection with Spark. Information Fusion, 2018: 182–194.
[7] Carcillo F, Le Borgne Y A, Caelen O, et al. Streaming active learning strategies for real-life credit card fraud detection: assessment and visualization. International Journal of Data Science and Analytics, 2018(4): 285–300.
[8] Bhattacharyya S, Jha S, Tharakunnel K, et al. Data mining for credit card fraud: A comparative study. Decision Support Systems, 2011(3): 602–613.
[9] Davis J, Goadrich M. The relationship between Precision-Recall and ROC curves. Proc. 23rd International Conference on Machine Learning (ICML), 2006: 233–240.
[10] Breiman L. Random forests. Machine Learning, 2001(1): 5–32.
[11] Pedregosa F, Varoquaux G, Gramfort A, et al. Scikit-learn: Machine learning in Python. Journal of Machine Learning Research, 2011: 2825–2830.
[12] Ke G, Meng Q, Finley T, et al. LightGBM: A highly efficient gradient boosting decision tree. Advances in Neural Information Processing Systems (NeurIPS), 2017: 3146–3154.
[13] Bottou L. Large-scale machine learning with stochastic gradient descent. Proc. COMPSTAT 2010, 2010: 177–186.