Topological representation learning for e-commerce shopping behaviors

Yankai Chen, Tuan Truong, Xin Shen, Ming Wang, Jin Li, Jim Chan, Irwin King

Abstract: Learning compact representation from customer shopping behaviors is at the core of web-scale E-commerce recommender systems. At Amazon, we put great efforts into learning embedding of customer engagements in order to fuel multiple downstream tasks for better recommendation services. In this work, we define the notion of shopping trajectory that consists of customer interactions at the categorical level of products, then construct an end-to-end model namely C-STAR which is capable of learning rich embedding for representing the variable-length customer trajectory. C-STAR explicitly captures the trajectory distribution similarity and trajectory topological semantics, providing a coarse-to-fine trajectory representation learning paradigm both structurally and semantically. We evaluate the model on Amazon proprietary data as well as four public datasets, where the learned embeddings have shown to be effective for customer-centric tasks including customer segmentation and shopping trajectory completion.