Learned Temporal Aggregations for Fraud Classification on E-Commerce Platforms

Fraud and other types of adversarial behavior are serious problems on customer-to-customer (C2C) e-commerce platforms, where harmful behaviors by bad actors erode user trust and safety. Many modern e-commerce integrity systems utilize machine learning (ML) to detect fraud and bad actors. We discuss the practical problems faced by integrity systems which utilize data associated with user interactions with the platform. Specifically, we focus on the challenge of representing the user interaction events, and aggregating their features. We compare the performance of two paradigms to handle the feature temporality when training the ML models: hand-engineered temporal aggregation and a learned aggregation using a sequence encoder. We show that a model which learns a time-aggregation using a sequence encoder outperforms models trained on handcrafted aggregations on the fraud classification task with a real-world dataset.