A Survey Of Collaborative Filtering Based On Transfer Learning

Recommender system is the most effective approach to solve the informationoverload problem, which applys the history of users’ rating behavior to providepersonalized recommendation for human. As the core of recommender system,recommendation algorithm is considered to determine the quality of recommendation.Collaborative filtering is one of the most widespread and successful techniques forrecommender system. However, collaborative filtering is facing some challenges withthe growing of users and items, such as data sparsity, scalability and so on. In order toalleviate the data sparsity problem, transfer learning methors extract the knowledgefrom auxiliary domain that has more training data and applies the knowledge to targetdomain, which will help improving the recommender accuracy for the latter. Ourwork studies the collaborative algorithms from perspective of transfer learning, andwe propose three transfer learning model to adapt to different situations of datadistribution in auxiliary domain and target domain.(1) In order to improve the accuracy of user similarity for traditional collabo-rative filtering algorithm, we construct a collaborative filtering model based on usersimilarity transfer, which uses the user similarity from auxiliary domain to help thelearning of user similarity in target domain. Furthermore, main fators that may affectthe model balance parameter have been analysis in the paper, and we use a featuresubspace distance to estimate the parameter approximately, which will make therecommendation system more intelligent. Through the experiment on real world data,the results show that our model can calculate the user similarity accurately in targetdomain, and improve the quality of recommendation.(2) Base on matrix factorization model, we propose a user feature transfer ofcollaborative filtering recommender model to help the learning of user feature intarget domain. While for the task of auxiliary domain, we construct user preferencestructure by adopting the SOFT-IMPUTE algorithm to solve a nuclear norm regulari-zed least-squares problem. In addition, a fast convergence Wiberg algorithm isdesigned for the model to get the optimal solution. By the experiment on the real data,the model can effectively transfer the user feature of source domain, and improve thequality of recommender system for target domain.(3) To overcome the limitation that user features or items features learned fromdifferent domain must have the same dimensisons, we propose a transfer learningmethod for collaborative filtering, called Feature Subspace Transfer. In the model, wemeasure the similarity between different domains’ user feature by the projectiondistance from target domain’s user feature to the subspace spanned by auxiliary domain’s user feature. Furthermore, an iterative algorithm is also proposed forsolving the objective optimization problem. Numerical experiments on real worlddata show the improvement of our model on alleviation the sparsity problem.