Real-time Fraud Detection using Deep Learning
summary
Claims an autoencoder reconstruction approach plus SHAP explanations detects anomalies and traces root input variables; model trains on normal transactions so high reconstruction loss marks fraud. Pipeline processes featureized records with up to ~500 flags per transaction, initially batching 30 minutes of data in ~10 minutes then moving to Spark Streaming for near real‑time scoring. Implementation at PayPal Chennai replaces brittle rule sets and produces actionable root‑cause scores for domain experts to review.
AI-written summary · about
chapters (5)
- 0:00Introduction
- 1:46Limitations of rule-based systems
- 8:35Proposed architecture overview
- 13:13Autoencoder model basics
- 16:50Model tuning factors
related talks
52mModern Fraud Prevention Using Deep Learning
28mAnomaly Detection using Autoencoders
25mUnsupervised shallow learning for fraud detection on marketplaces
29mDeep Learning Applications to Online Payment Fraud Detection
40mFraud detection without labeled data
37mThe Role of Machine Learning in Today’s Payment Ecosystems
10mReal-Time Fraud Prevention at Scale
16mLeveraging ML to Create a Reliable Anti-Fraud System