Collaborative Filtering Study Based On Electrical Resistance Network And Sparse Data Prediction

There are one million digital books collected in China-America Digital Academic Library (CADAL) so that our users often confront with information overload problems in their browsing process. We provide several personalized services in CADAL portal, among which personalized recommendation service can help users overcome the information overload problem, which really saves their time and labor cost.Over the past decade, collaborative filtering algorithms applied to many popular recommender systems have made remarkable progresses. However, there are two major problems: (1) the ways of similarity computation among users or items are just heuristic, e.g. most of them are vector calculation. (2) Given the sparse rating data, the memory-based approaches exploited by many recommender systems perform not well. Especially in the beginning phase of the systems, there will exist cold-start problem.The work of this paper can be summarized as follows: (1) we adopt a novel similarity computation algorithm based on electrical resistance network. Users and items can be viewed as the nodes of the electrical resistance network, whose edges correspond to the ratings, and the rating values are regarded as the conductance. Therefore, we can compute the similarity among users or items by solving linear equations as in the physical electrical resistance network. (2) In order to alleviate the problem of data sparsity, we adopt a sparse data prediction algorithm which combines user and item information, and predicts the sparse data of the user-item matrix optionally in terms of our criterion. Finally, we predict items' rating values possibly given by users based on the extended user-item matrix. (3) In the end, we implement Electrical Resistance Network Model for Collaborative Filtering (ERCF). In terms of the little number of ratings collected in CADAL portal, we test experiments on the notable MovieLens dataset. The results show that ERCF outperforms other collaborative filtering algorithms we test, and prediction of sparse rating information optionally can improve the accuracy of recommender algorithm.