SLLIM-Rank: A Multi-Stage Item-to-Item Recommendation Model using Learning-to-Rank

Item-to-item recommendations are crucial for user content discovery and engagement on online platforms, often showcased in prominent areas like "You May Also Like." These models typically leverage metadata and user engagement data to generate recommendations; however, data sparsity presents challenges, particularly when new movies or shows are released, limiting the ability to provide optimal recommendations early on. Additionally, as users access content across various devices with different screen sizes, it is essential to optimize the ranking of recommendations to ensure the most relevant items appear at the top. Finally, with platforms serving millions of users and an ever-changing inventory of items, scalable methodologies are necessary to effectively address these challenges. In this paper, we propose a scalable multi-stage item-to-item recommendations model called SLLIM-Rank: Similarity with Large Language Improved Model using Learning-to-Rank. The approach utilizes (a) temporal and contextual features to capture dynamic trends in item similarity, (b) a Learning-to-Rank model to prioritize items based on implicit user feedback and, (c) large language models (LLMs) to generate supplementary metadata for catalog items. We discuss effective strategies for offline evaluation of the model. Additionally, these offline findings lead to substantial improvements in key engagement metrics on a content streaming platform, specially improving the quality of cold item recommendations, demonstrating the high effectiveness of our approach in a real-world context.