Enabling Low Latency Fraud Detection with Real-Time Feature Engineering

Identifying fraudulent credit card transactions using a machine learning model relies on enrichment of those transactions so that they can be scored reliably. But how can complex stateful features such as average transaction amounts per customer be computed quickly enough? This is solved by building a real-time feature pipeline, using stream processing to compute the inputs for the machine learning model and storing them in a feature store, ready for low latency retrieval. Building real-time feature pipelines used to require complex frameworks, many of which wrap Python in another language like Java, making them notoriously difficult to debug. In this session we will dive into feature computation with a simpler native Python approach using the Quix Streams open source library.