Hybrid-Based Item Embedding for Cold-start Recommendations with SBERT

Abstract

Collaborative filtering, a primary method in information recommendation based on user behavior history, suffers from the cold start problem, which is the inability to handle items with few occurrences. As a solution, the introduction of hybrid-based filtering, which utilizes item information in addition to behavior history, has been suggested. However, when the behavior history is session information with anonymized user information, how to combine it with item information to create an embedding representation has not been studied. In this research, we propose a hybrid-based item embedding method that addresses the cold start problem when the behavior history is session information. Specifically, we consider the result of encoding the item title with SBERT as the embedding representation and fine-tuning SBERT with a triplet loss function using session information as positive examples. Experiments comparing the accuracy of this method revealed improvements in accuracy due to fine-tuning and that it outperforms existing content-based baseline methods.