Multi-tasking fraud detection From Trees to MLPs
summary
Monzo trains multi‑task neural nets in PyTorch alongside LightGBM/XGBoost baselines and finds statistically significant uplift in recall, with the best scenario showing 30% relative improvement and much better detection of held‑out fraud subtypes. Transaction embeddings from the shared layers reveal both general and specific fraud signals, and feeding those embeddings into LightGBM yields the top tradeoff of accuracy and speed, meeting the 950 ms latency constraint for real‑time deployment.
AI-written summary · about
chapters (9)
- 0:00Fraud detection at Monzo
- 3:54Why fraud detection is critical
- 8:25Limitations of tree models
- 10:55Multitask machine learning for fraud
- 15:01Testing on card fraud
- 17:47Results and performance
- 21:44Key findings and impact
- 22:24Deep learning with latency
- 23:28Summary
related talks
12mCloud-Native Friendly Fraud Detection
16mLeveraging ML to Create a Reliable Anti-Fraud System
52mModern Fraud Prevention Using Deep Learning
25mUnsupervised shallow learning for fraud detection on marketplaces
14mCloud-Native AI for Fraud Prevention with Bayesian Networks
13mHarnessing Machine Learning to Combat E-commerce Fraud
37mThe Role of Machine Learning in Today’s Payment Ecosystems
26mAI-Powered Fraud Detection in Cloud-Native Digital Payments