Learned sparse retrieval is a common technique for building recommender systems, but many current approaches, such as STAR, represent a user by concatenating their entire interaction history into a single query string. A new paper on arXiv argues that this design creates two bottlenecks for news recommendation. The concatenation collapses the structure of the user's history, losing information that could be useful for predicting what they want to read next.
The paper, titled 'LSNR: Learned Sparse News Recommendations with Attentive User Modeling' (arXiv:2610.23510), proposes a method that uses attentive user modeling instead. By avoiding the single-query concatenation, LSNR aims to preserve more of the user's interaction patterns. The abstract is truncated, so the full details of the method and any experimental results are not available from the source.