Learning Collective Embeddings for Item Cold-start Recommendations from Sentence Transformers

Traditional Collaborative Filtering using Matrix Factorisation predicts user preference for every item in a catalogue, only accepting user and item identifiers as input. Without access to side information about the user and item, or adequate user-item interactions, it runs into a situation where it does not have enough information about a user or item to produce decent recommendations for either, a situation called cold start. One of the ways to alleviate this problem is to include side information i.e. metadata about the user or items. In this work, we propose to jointly factorise the partially-observed preference matrix and text-rich item side information. We represent the side information as a pairwise similarity matrix of embeddings derived from Sentence Transformers. This representation ensures we capture the relationships across every item. Our aim is to find item factors such that knowledge about the item metadata alleviates item cold start. To do that, we align the pairwise relationship between latent item factors from the decomposition of the user interaction matrix with the pairwise similarity matrix of sentence embedding by taking the Frobenius norm of their differences. This alignment penalty preserves global structure during joint factorisation, capturing the semantic relationship across all items. The results obtained from our experiments show that our approach effectively exploits the rich semantic representation of text derived from Sentence Transformers to improve upon existing methods that relied on a sparse representation of side information and Collaborative Filtering.