Study On Bipartite Network In Collaborative Filtering Recommender System

Open Access
Yao, Luqi
Graduate Program:
Industrial Engineering
Master of Science
Document Type:
Master Thesis
Date of Defense:
Committee Members:
  • Dennis Kon Jin Lin, Thesis Advisor
  • recommender system; bipartite network; collaborative filtering;
Recommender system is increasingly popular in recent years, scientists came up with plenty of recommendation algorithms and never stop trying to make the recommendation more accurate. Although the relationship within recommender systems could usually be represented as a network which is few people tried to build recommender systems based on the properties of the network. This might be a potential area that is underdeveloped. In this thesis, the projection of a weighted bipartite graph used in the collaborative filtering recommender system will be presented. Based on previous work in the related field, the similarity function of general collaborative filtering is redefined using the resource allocation process on the weighted bipartite graph. The process of resource allocation is implemented by a two-step random walk which calculates the recommendation power between each user. The recommendation power within the set of users is used to generate the specific similarity function for collaborative filtering. The proposed new method is used in the MoiveLen dataset which is a benchmark dataset and its performance is compared with four widely used methods: collaborative filtering using Pearson correlation, collaborative filtering using cosine similarity, user-mean, and item-mean by comparing the value of MAE and RMSE. However, the results is not as good as expected. The performance of projected bipartite method is not outstanding. It needs to be improved in the future.