Utilizing Embeddings to Learn a Universal Customer Behavior Representation in E-Commerce

A decoupled three-stage approach is introduced, combining self-supervised learning of customer embeddings from behavioral data with flexible downstream models for predicting customer intentions.

E-commerce operates in a highly dynamic and competitive environment, where customer satisfaction is key to success. Delivering personalized experiences at scale requires systems capable of reliably modeling individual customer behavior while respecting privacy and data protection constraints such as the GDPR. This book proposes a universal, privacy-compliant customer representation that is task-agnostic and incrementally adaptable. A decoupled three-stage approach is introduced, combining self-supervised learning of customer embeddings from behavioral data with flexible downstream models for predicting customer intentions. Temporal extensions improve performance, particularly under sparse information conditions, while lifelong learning enables dynamic adaptation to new interactions and evolving product spaces without full retraining. Comprehensive experiments across multiple real-world e-commerce datasets demonstrate consistent performance improvements over state-of-the-art baselines. By decoupling personalization from personal data, this work offers a scalable and privacy-preserving foundation for next-generation personalization systems.