Fantastic Embeddings and How to Align Them: Zero-Shot Inference in a\n Multi-Shop Scenario

Abstract

This paper addresses the challenge of leveraging multiple embedding spaces\nfor multi-shop personalization, proving that zero-shot inference is possible by\ntransferring shopping intent from one website to another without manual\nintervention. We detail a machine learning pipeline to train and optimize\nembeddings within shops first, and support the quantitative findings with\nadditional qualitative insights. We then turn to the harder task of using\nlearned embeddings across shops: if products from different shops live in the\nsame vector space, user intent - as represented by regions in this space - can\nthen be transferred in a zero-shot fashion across websites. We propose and\nbenchmark unsupervised and supervised methods to "travel" between embedding\nspaces, each with its own assumptions on data quantity and quality. We show\nthat zero-shot personalization is indeed possible at scale by testing the\nshared embedding space with two downstream tasks, event prediction and\ntype-ahead suggestions. Finally, we curate a cross-shop anonymized embeddings\ndataset to foster an inclusive discussion of this important business scenario.\n