ИСТИНА

Data mining for recommender systems has gained a lot of interest in the recent years. ”ECML/PKDD Discovery Challenge 2011” was organized to improve current recommender system of the VideoLectures.Net website. Two main tasks of the challenge simulate new-user and new-item recommendation (cold-start mode) and clickstream based recommendation (n ormal mode). This paper provides detailed descriptions of two simple algorit hms which were very successful in the both tasks. The main idea of the algorithms is construction of a linear combination equal to a vector of estimations of lectures popularity after viewing a certain lecture (or lectures). Each addend in the combination describes similarity of lectures using the part of the data. The algorithms are improved by transforming the combination to non-linear function. Lectures with the highest estimations of popularity are recommended to users.