Detecting fraud faster with Amazon SageMaker, featuring Coinbase
summary
Coinbase describes a production fraud stack built on Amazon SageMaker that achieves a 50% reduction in fraud year over year, uses ensemble models like CatBoost + transformers (PyTorch) for user, transfer, credit and account takeover risk, and enforces an 80 ms SLA end to end. Training and inference use Bring‑Your‑Own‑Container, SageMaker endpoints, Autopilot and built‑in algorithms; a streaming feature store computes features with Flink over Kinesis and serves via DynamoDB. AWS provides two launchable templates: a tabular supervised+unsupervised pipeline and a graph GNN pipeline using Deep Graph Library.
AI-written summary · about
chapters (15)
- 0:00Intro
- 1:24Amazon SageMaker solutions for fraud detection
- 2:30Typical approach to combat fraud: Rules based systems
- 3:30Rules based fraud detection - Limitations
- 4:12Machine learning for fraud detection
- 5:59Building Fraud detection models on Amazon SageMaker
- 8:40Amazon SageMaker features - Model development
- 12:52Coinbase fraud prevention ML models
- 13:52Model development lifecycle
- 17:00Prototype using Jupyter Notebooks on Amazon SageMaker
- 18:44Training models in production using Amazon SageMaker
- 19:18Checking in models and achieving consensus
- 20:30Deploying real time models
- 22:41Meeting Coinbase ML Fraud Detection Platform requirements
- 25:10SageMaker solution: Fraud detection using Graph Neural Networks
related talks
39mFraud Detection: Using ML to Identify and Manage Fraudulent Activities
19mUsing Amazon SageMaker to develop and deploy ML models
19mDrive transformation through machine learning
43mMachine Learning Ops: DevOps for Data Science
37mStreamlining Machine Learning Operations
51mProfundizando en Aprendizaje Automático para Desarrolladores
19mDeveloping large-scale machine learning algorithms on Amazon SageMaker
37mStreamlining Machine Learning Operations