Harnessing Temporal Dynamics and Content: An Ensemble of Gradient Boosting Machines for News Recommendation

This paper presents our approach for Team Tom3TK that achieved third place in the ACM RecSys Challenge 2024, organized by Ekstra Bladet. The challenge centered on large scale news recommendations with over 380 million impression logs and aimed to predict which article a user will click on from a series of articles displayed in a specific session. Our approach utilizes a LightGBM ensemble model, integrating article timeliness features with content-based recommendations from historical implicit feedback and employs advanced multilingual embedding models. Experiments revealed the significant impact of timeliness and content-based features, the importance of suitable embedding models, and the phenomenon of performance plateauing with the expansion of training samples.