Evaluating Ensemble Methods for News Recommender Systems

This paper investigates the potential of ensemble methods in News Recommender Systems (NRS) to enhance performance. It evaluates diverse algorithms, such as BERT content-based and LSTUR collaborative filtering, on the Microsoft News Dataset (MIND), finding that combining diverse base learners can lead to up to a 5% improvement. However, combining non-distinct methods does not result in improvement. The study highlights the significance of selecting diverse algorithms for optimal ensemble results and suggests the need for further research in this area, while also acknowledging the challenges of cold-start problem and the evolving nature of data-driven methods like the poly attention mechanism in Miner. Performance metrics like AUC and MRR are used to assess the models, with some studies indicating that combining the best content-based and collaborative filtering methods can lead to substantial improvements. Ethical considerations, such as avoiding echo chambers and biases, are also mentioned as important for future research.