Research On Collaborative Recommendation Algorithm For Explicit Feedback And Implicit Feedback

With the fast development of e-commerce services, users' online behaviors such as numeric ratings, bookmarks(e.g., adding to cart) and browsing records are regarded as one of the most valuable sources of information to learn users' interests.However, most previous research only focus on one single type of users' behaviors,such as explicit feedbacks of numeric ratings or implicit feedbacks of browsing records. Few research have studied both explicit feedbacks and implicit feedbacks in one algorithm. In this thesis, we study those two types of feedbacks simultaneously,for two different recommendation tasks, i.e., collaborative regression and collaborative ranking. In order to integrate those two types of feedbacks, there are some challenges such as heterogeneity of users' feedbacks and uncertainty of implicit feedbacks. As a response, we design a novel self-transfer learning(sTL) algorithm for the collaborative regression task, and a novel staged matrix factorization(SMF)algorithm for the collaborative ranking algorithm.In order to study the effectiveness of our proposed sTL algorithm, we conduct extensive experiments on three large data sets. The experimental results show that our sTL algorithm can get much better recommendation performance than other recommendation algorithms such as SVD and SVD++. Similarly, we study the performance of our SMF algorithm on three large data sets, and find that our SMF can make better use of explicit feedbacks and implicit feedbacks as compared with other recommendation algorithms.As far as we know, in most recommendations systems, there are usually both explicit feedbacks and implicit feedbacks, and there are few only one single type of users' feedbacks. Hence, our proposed sTL and SMF algorithms have values for real e-commerce services.