Research On Methods Of Cross-Domain Recommendation With Transfer Learning

With the development of computer technology and the expansion of the network, people has been surrounded by the sea of information. Recommendation system was advanced under this background. Recommendation system is a hot research field, and it has been successfully applied on E-commerce sites. It recommended contents for users which they are interested in numerous information.To make the users get the information quickly. At present, most of the recommendation system is based on collaborative filtering recommendation technology. But there are exist two main problems: data sparsity problem and cold start problem.At present, some models based on transfer learning have been proposed to alleviate the data sparsity problem and cold start problem. But they do not take the diversity among the related domains into account. The diversity might clean the advantages of common pattern, which may result in bad performance. For the problem, we proposed two new cross-domain recommendation methods with transfer learning:1) This thesis proposes a transfer learning method based on tags. Most of recommendation have tags which used to describe the features of users and items.We can exploit the tags to connect the two domains. The algorithm is mainly divided into three steps: Firstly, we classify the users from auxiliary domain by non-negative orthogonal matrix tri-factorization, so we can get the different type of user; Then we train the BP neural network through the feature tags of different type of users, and the BP neural network can judge the type of user by their features. At last, we get all the types of the users from target domain by the trained neural network, then predict the missing value in target matrix by the average value of the rating which the same type of user rating on the same item.2) This thesis proposes a cluster-level based latent factor method. This method can not only learn the common rating pattern shared across domains with the flexibility in the controlling the optimal level of sharing, but also learn the domainspecial rating patterns of users in each domain that involve the discriminative information propitious to performance improvement.