Development of a hybrid recommendation system

Abstract

The article discusses the solution to the problem of informing the user about the service that is most interesting to him at a given time. For this, an analysis of modern approaches to the construction of recommender systems has been carried out, their advantages and disadvantages have been identified. The main ones are the cold start problem, bad predictions for atypical users, computational resource consumption, trivial predictions. Further, the metrics for assessing the quality of such systems are considered. Several recommender systems have been developed with the following approaches: a system using collaborative filtering by users, a system using collaborative filtering by subjects, a system of categorical interests, and Slope one approach. To find similar users or items in a system using collaborative filtering, the k-nearest neighbors method was used. Based on the data obtained, it is proposed to build a hybrid recommender system that compensates for the shortcomings of each of the approaches considered. The personalized approach is based on a Bayesian rating. The non-personalized approach is based on Slope One algorithms and collaborative filtering by subject. In the case when the user has not rated any of the services, recommendations are built using a collaborative filtering approach by user, based on his views and purchases. The training and testing of algorithms for the organization of the built recommender systems is carried out. Thus, a recommendation system for services has been implemented, capable of generating recommendations for both registered and unregistered users. An analysis of the effectiveness of recommendation systems was carried out using the Amazon review data dataset. Based on the information received, the developed hybrid recommendation system leads to the best solution.