Applying hard and fuzzy K-modes clustering for dynamic Web recommendations

This paper proposes a dynamic Recommender System for the Web, which uses Entropy based Hard and Fuzzy K-modes algorithms to group items in clusters according to certain data features. Recommendations are then produced from those clusters that have their centers closer to the user's search preferences. The system starts operating with a learning session which allows the user to conduct various searches in order to identify her/his initial preferences. The ongoing searching behavior of the user is dynamically recorded and inserted in the recommendation engine, adjusting user preferences in real-time and enabling the system to maintain high levels of accuracy. The proposed approach is validated on a movie dataset containing information on stars, categories and production companies, with the user being able to search for an item based on one or more of these features. The results indicate successful performance for the two clustering algorithms, while a short comparison to variations of the KNN algorithm suggests superiority of the proposed approaches