Venkatesh Ramanathan, Paypal - Fraud Detection Using H2O’s Deep Learning
- Predictive AI
- H2O-3
Overview
PayPal data scientist Venkatesh Ramanathan presents an early production case study on using H2O's deep learning for fraud detection across PayPal's transaction, account, and network-level risk layers. He walks through the practical struggles of running deep learning at scale on a 10-million-row, 1,500-feature Hadoop dataset, including cluster memory limits, slow data import, and missing values, and how direct collaboration with H2O's engineering team resolved each one. He closes with tuning results across hidden layers, activation functions, and feature subsets, including a measured AUC improvement from deep learning's automatic feature derivation.
Key takeaways
- How PayPal structures fraud prevention across three layers: real-time transaction-level scoring, account-level abuse pattern monitoring, and network-level tracing of fraudster account clusters
- Why Venkatesh chose H2O's deep learning over other options: scalability, benchmarked performance, flexible deployment, and native integration with Hadoop and Spark
- The real infrastructure problems encountered running H2O at scale, including cluster memory constraints, slow single-threaded HDFS import, and models failing to train because 99% of rows had missing values
- How tuning hidden layers (settling on six), activation functions (rectifier with dropout performed best), and feature subsets affected model accuracy
- How deep learning's ability to derive nonlinear features on its own improved AUC by about 1% on a feature subset that lacked engineered nonlinear features
- How H2O's "replicate training data" flag trades network overhead for training speed by letting each cluster node train on only its local data slice, then averaging the resulting models
Start owning your AI