On a Novel Machine Learning Based Approach to Recommender Systems

A new approach for recommender systems design is proposed. The considered system should rely only on the anonymous receipts' data and information about products currently bought by a customer. The preference rating for an arbitrary product is calculated as a classification result of a combined feature description of the product and the currently bought ones. The corresponding product descriptions are formed by vectors of distances between the products and precalculated product clusters obtained by applying hierarchical clustering technique to large binary product to receipt relevance matrix. The proposed method is compared with two other techniques in experiments with real retail data. The first one evaluates preference rating simply as a product sales rate. The second technique uses association rules combinations. The better performance of the proposed approach was observed.