Systems and methods for universal item learning in item recommendation

Abstract

Embodiments described herein provide a universal item learning framework that generates universal item embeddings for zero-shot items. Specifically, the universal item learning framework performs generic features extraction of items and product knowledge characterization based on a product knowledge graph (PKG) to generate embeddings of input items. A pretrained language model (PLM) may be adopted to extract features from generic item side information, such as titles, descriptions, etc., of an item. A PKG may be constructed to represent recommendation-oriented knowledge, which comprise a plurality of nodes representing items and a plurality of edges connecting nodes represent different relations between items. As those relations in PKG are usually retrieved from user-item interactions, the PKG adapts the universal representation for recommendation with knowledge of user-item interactions.