Search-based Recommendation: the Case for Difficult Predictions

Recommender systems have achieved impressive results on benchmark datasets. However, the numbers are often influenced by assumptions made on the data and evaluation mode. This work questions and revises these assumptions, to study and improve the quality, particularly for the difficult case of search-based recommendations. Users start with a personally liked item as a query and look for similar items that match their tastes. User satisfaction requires discovering truly unknown items: new authors of books rather than merely more books of known writers. We propose a unified system architecture that combines interaction-based and content-based signals and leverages language models for Transformer-powered predictions. We present new techniques for selecting negative training samples, and investigate their performance in the underexplored search-based evaluation mode.