Data imputation using a trust network for recommendation

Abstract

Recommendation methods suffer from the data sparsity and cold-start user problems, often resulting in low accuracy. To address these problems, we propose a novel imputation method, which effectively densifies a rating matrix by filling unevaluated ratings with probable values. In our method, we use a trust network to estimate the unevaluated ratings accurately. We conduct experiments on the Epinions dataset and demonstrate that our method helps provide better recommendation accuracy than previous methods, especially for cold-start users.