Дослідження методу взаємного злиття рангів у методах колаборативної фільтрації рекомендаційних систем на основі неявного зворотного зв’язку

Abstract

The study investigates the application of the RRF method in CF tasks of recommendation systems for combining the results of different recommendation algorithms. The RRF method is used within the late fusion strategy for the formation of a single ordered list of recommendations based on the results of the following algorithms: ALS, BPR, BiVAE, LightGCN, NCF, and SAR. The experimental evaluation covers fifteen pairwise combinations of algorithms, different in their construction principles, tested on the MovieLens 100K dataset and two subsets of datasets from the Amazon Reviews'23 collection (Video Games and Movies and TV) converted to the implicit feedback format. The experimental results demonstrate that applying the RRF method in the late fusion strategy mostly improves the metrics for the formation and ordering of recommendation lists compared with using individual recommendation algorithms.