Measure prediction capability of data for collaborative filtering

Collaborative filtering (CF) approaches have been widely been employed in e-commerce to help users find items they like. Whereas most of existing work focuses on improving algorithmic performance, it is important to know whether the recommendation for users and items can be trustworthy. In this paper, we propose a metric, “relatedness,” to measure the potential that a user’s preference on an item can be accurately predicted. The relatedness of a user–item pair is determined by a community which consists of users and items most related to the pair. The relatedness is computed by solving a constrained \(\ell _{1}^{}\)-regularized least square problem with a generalized homotopy algorithm, and we design the homotopy-based community search algorithm to identify the community by alternately selecting the most related users and items. As an application of the relatedness metric, we develop the data-oriented combination (DOC) method for recommender systems by integrating a group of benchmark CF methods based on the relatedness of user–item pairs. In experimental studies, we examine the effectiveness of the relatedness metric and validate the performance of the DOC method by comparing it with benchmark methods.