Similarity Learning Algorithms For One-Class Feedbacks In Recommender Systems

Recently, item-based similarity learning for recommendation with one-class feedbacks has attracted lots of attentions. Specifically, it aims to generate a ranked top-N list of items for each user via learning item-item similarities from users' one-class feedbacks, such as purchase behaviors with positive preferences and/or click behaviors with uncertain preferences. In many real situations, the purchase behaviors are usually very few, which makes the user-item matrix rather sparse. The sparseness problem will then hurt the item-item similarity learning and recommendation performance.Most recent works only exploit the purchase data, while the more abundant click or browsing data is rarely modeled simultaneously. We believe that incorporating the click data into a recommendation model has a potential of enhancing the recommendation performance. Furthermore, we find that there are at least two major approaches to estimate the item-item similarity, including(i) similarity calculation via cosine similarity or Jaccard index as usually used in traditional item-based collaborative filtering algorithms, and(ii) similarity learning via some machine learning techniques such as the seminal work FISM(factored item similarity model). Based on the perspective of similarity calculation and similarity learning, we surprisingly find that no previous work has studied the complementarity of those two approaches yet.In this thesis, we mainly generalize the famous FISM model, and propose two specific novel algorithms for recommendation with one-class feedbacks. Firstly, we study a new recommendation problem called heterogeneous one-class collaborative filtering(HOCCF) including both purchase data and click data, and design a novel item-item similarity learning algorithm, i.e., transfer via joint similarity learning(TJSL). TJSL jointly learns a similarity between a candidate item and a preferred item, and a similarity between a candidate item and an identified likely-to-prefer item. Experimental results show that our TJSL can achieve better performance than FISM with purchase data only. Secondly, we study the classical one-class collaborative filtering(OCCF) problem including one type of feedbacks only, and develop a novel mixed similarity learning algorithm, i.e., pairwise factored mixed similarity model(P-FMSM). P-FMSM aims to combine the merits of similarity calculation and similarity learning in one single framework, and is also very efficient via similarity factorization. Empirical studies show that our P-FMSM can generate more accurate recommendations than FISM.