Computing User Similarity by Combining SimRank++ and Cosine Similarities to Improve Collaborative Filtering

Abstract

This paper addresses the sparsity problem in collaborative filtering (CF) by developing an aggregated useruser similarity measure suitable for the user-based CF model. The aggregated similarity measure is a weighted aggregation of the SimRank++ similarity on the user-item bipartite graph and the cosine similarity of the Linked Open Data (LOD)-based user profiles derived from both the rating data and the items' descriptive attributes found from LOD resources. To validate the effectiveness of the aggregated similarity and evaluate the accuracy of rating predictions with the user-based CF method, comparative experiments between four similarity measures, the Pearson correlation coefficient, the SimRank++ similarity, the cosine similarity and the aggregated similarity, were conducted on the MovieLens 100k dataset and DBpedia. The experimental results indicate that the proposed aggregated similarity measure overall outperforms the other three similarity measures in terms of both Root Mean Squared Error (RMSE) and Mean Absolute Error (MAE), especially in the cases of 30-100 nearest neighbors.