Designing Hybrid Recommender Systems
Anna V. Zagranovskaia, Dmitry Mitura · 2021
The purpose of the study is to analyze the existing methods of building hybrid recommender systems and develop the most accurate and adequate movie recommender system of that type. The paper addresses basic algorithms for designing recommender systems (content-based, collaborative filtering, hybrids), and corresponding similarity metrics (cosine distance, Pearson correlation coefficient). The interest in hybrid systems is due to the fact that combining different modelling methods can improve not only the accuracy of rating predictions but also other characteristics of a recommender system, such as novelty, serendipity, diversity. In the course of the study, three approaches were used: content-based, collaborative and hybrid ones. As a result, two best models were chosen based on the first two approaches, with reference to which a hybrid model was built, whose quality was evaluated through a survey of potential users of the system. This made it possible to cover a whole range of evaluation criteria for recommender systems and understand corresponding user requirements. Trust to the system turned out to be the most important one. In other words, it means the extent to which the user is confident that he or she will actually like the recommendations suggested. Recommender systems were tested using open source data of the MovieLens dataset. This makes it possible to verify the results and improve the models suggested.