RECOMMENDER SYSTEMS: PREDICTING WITH MACHINE LEARNING BASED ON POPULATION-BASED ALGORITHM

Abstract

The article considers the formulation of the forecasting problem as well as such problems of recommender systems as data sparsity, cold start, scalability, synonymy, fraud, diversity, white crows. Combining the results of collaborative and content filtering gives us two possibilities. On the one hand, to weigh the results according to the content data. On the other hand, to shift these weights towards collaborative filtering as soon as data about a particular user appears. In turn, this improves the accuracy of the recommendations. The authors propose a hybrid model of a recommender system. Such a system includes the characteristics of collaborative and content filtering both. Also, the population-based algorithm for filtering and the architecture of a recommendation system based on it are described in the article. The algorithm consists of the following steps: study the search space; synthesis of solutions, i.e. points of this space; request quality assessment decisions or “fitness”; using it to make “natural selection”. Here we see the learning process about which areas of the search space contain the best solutions. The population of user “characteristics” encoded in the population-based algorithm supports a variety of input data in a hybrid model. The authors propose a coding structure for decisions in a population-based algorithm using the example of a recommender movie viewing system. Drift analysis evaluates the polynomial complexity of the algorithm. The authors demonstrate the results of experimental studies on an array of benchmarks. We also present an assessment of filtration efficiency based on a hybrid model and a population-based algorithm in comparison with the traditional method of collaborative filtering using the Pearson correlation coefficient. We can see that the prediction accuracy of the population-based algorithm is higher than that of the Pearson algorithm.