Pairwise Preferences Based Recommendation Techniques

Recommender Systems (RSs) are software tools and techniques which learn user's interest and recommend items that he/she will find most useful among the available ones. Most of the present research and application of Recommender Systems use user preference data that are in the form of ratings. However, this type of preferences has few disadvantages. For instance, if most of the user rated items are 5 stars, then it is difficult to understand which item the user prefers among them. In this research work, we focus on pairwise preferences as an alternative way for modeling user preferences and compute recommendations. In our scenario, users compare a set of item pairs, indicating which item in each pair, and to what extent, is preferred. Actually, in this thesis, we aim at combining both ratings (absolute feedback) and pairwise preferences (relative feedback) in order to make the best use of this mixed preference data. Furthermore, we aim at identifying specific conditions/situations where pairwise preferences elicitation is meaningful and beneficial. Having the goal of incorporating pairwise preferences along with ratings in RS, in this thesis, we have designed recommendation techniques using pairwise preference and ratings. In particular, our main contributions include the followings: (i) new recommendation techniques to be used in a Matrix Factorization (MF) and Nearest Neighbor (NN) based pair scores prediction and ranking methods; (ii) in order to make the proposed recommender system more flexible and more accurate, we extended our proposed recommendation model based on pairwise preferences expressed over items so that it can also exploit preferences expressed over features; (iii) we propose a novel GUI for a mobile application to elicit pairwise preferences in a mobile environment and a complementary pair score elicitation algorithm for identifying which item pairs bring the most useful information about the user?s preferences, and hence should be asked from the user to compare; and finally (iv) we perform a live user study to understand the best scenario for a systems to ask which type of preferences (ratings or pairwise preferences) from users and collect their useful preferences on items. To evaluate the proposed solutions, we have conducted both offline and online experiments, using a variety of testing methods,  datasets, and a broad range of evaluation metrics. Our results reveal that our recommendation techniques can be effectively used to build pairwise preference based RS and compute accurate recommendations. Moreover, our techniques can also be used to incorporate feature based preferences for a RS and have shown that such preferences can help to improve the system performance, especially in the cold start phase.  Finally, by conducting an A/B test,  we identified specific conditions and situations where pairwise preferences elicitation is meaningful and beneficial, and we have shown that, when the user is searching for a specific recommendation, RSs with pairwise preferences can perform equally or better than state of the art rating based solutions.