Binary-data multi-criteria recommender systems based on Naive Bayes classifier

Recommender systems are specialized in suggesting appropriate items to users with respect to their personal characteristics and past preferences without requiring any effort of users. It might be more efficient to collect preferences of users based on multiple sub-criteria of corresponding product or service. For this purpose, researchers have proposed multi-criteria recommender systems that are convenient for more accurate and effective evaluation of items. In such systems, it might be preferable to collect binary ratings instead of numerical ones due to large number of sub-criteria. Naïve Bayes classifier is used for collaborative filtering purposes in single-criterion based recommender systems utilizing binary data. However, there is a gap in the literature in terms of a similar multi-criteria system. In this thesis, applicability of multi-criteria recommender systems based on binary data is investigated. Firstly, recommendations for users on overall preference criterion are produced employing naïve Bayes classifier. In order to improve quality of recommendations, user and item based similarity models are proposed enabling formation of more successful neighborhoods. Such models are further improved by integrating concordance measure between overall preference and sub-criteria ratings. Concordance measure provides the opportunity to calculate more personalized similarities among users. Finally, a hybrid model is proposed facilitating employing user and item based models together and statistically significantly improving quality of estimated binary referrals.