Leveraging LightGBM Ranker for Efficient Large-Scale News Recommendation Systems

This study addresses the news recommendation task presented by Ekstra Bladet in the ACM RecSys Challenge 2024. The task aims to predict which articles users are likely to click on from a list of candidate articles, by leveraging users’ browsing history, personal information, article details, and session information. Our approach is centered around a LightGBM Ranker model. Various features were used, including user, article, and session information, as well as their interactions. Additionally, embeddings were created from users’ browsing history and news article texts, and their cosine similarities were used as additional features. Appropriate validating methods using time series were explored, and effective data sampling and ensemble methods were also proposed to fit the data within limited memory. Finally, the final model was created by performing a weighted ensemble using multiple periods and random seeds. This method achieved high performance with AUC of 0.8169. As a result, an 8th place finish was achieved among around 200 participating teams. The code is available at https://github.com/tetsuro731/RecSys-Challenge-2024-tetsuro731.