The ongoing exponential growth of the Internet has promoted an information explosion, which subjects us to information overload: there are too much data to find the most relevant information. Therefore, information filtering is urgently required and the most important tools include search engines and recommender systems. Search engines provide keyword-based undifferentiated service, which means people will receive the same results when using the same keywords. Recommender systems provide personalized information filtering service initiatively, and are widely used in e-commerce and social networks. Collaborative Filtering(CF) methods mine user preferences from historical records of user behaviors rather than analyzing the content, and thus are the most popular in recommender algorithms. However, there are some drawbacks in the research of CF methods.(1) Stochastic Gradient Descent(SGD) is usually used to train Matrix Factorization(MF) models, but the error descent slows down with the increased iteration number, which may prolong the training procedure.(2) In the age of Big Data, e-commerce companies tend to cooperate by sharing datasets, but there is little related research.(3) Existing CF-based recommender algorithms are almost linear models, and some cannot even fuse social relations, which limits the ability to extract latent information from the rating matrix.In this paper, SGD was optimized to speed up the error descent by defining local structural information. In order to fuse different datasets, an MF method based on tag transfer learning was proposed. Finally, the Logistic function and social relations were used to develop nonlinear social MF models. The main work and achievements are as follows.1. An improved SGD method based on local structures was proposed. The difference matrix of the rating matrix was used to represent the local structures and to optimize SGD as a new objective function. The multi-objective optimization problem was solved hierarchically. An approximate optimal solution can be obtained on basis of the classic objective function, and a better solution around it can be found according to the new objective function. Experiments on two real-world datasets show that the improved algorithm has higher performance with fewer iterations.2. An MF model based on tag transfer learning was proposed. User features were represented by tag information in the auxiliary dataset, and were used to initialize the item features in the target dataset. User and item features were used to perform MF methods with target dataset after a smooth procedure. Experimental results show that the proposed algorithm predicts better and saves about half training time.3. Four nonlinear MF models based on the Logistic function were proposed to capture nonlinear interactions between latent factors. Both SGD and Markov Chain Monte Carlo(MCMC) methods were developed to train the proposed models. Experimental results prove the effectiveness of the Logistic function-based nonlinear models. The final model integrates the Logistic function and social relations into Bayesian Probabilistic Matrix Factorization model in a novel way, thereby significantly improving prediction accuracy at higher converging speed. |