Improving Prediction Accuracy in Neighborhood-Based Collaborative Filtering by Using Local Similarity

Neighborhood-based algorithms are some of the most promising memory-based collaborative filtering approaches for recommender systems. Many of these algorithms rely on a global similarity measure to select the most similar neighbors for rating prediction. However, these approaches may fail in capturing some meaningful relationships among users. In the real world, although users can show interest in a wide range of objects, they can express more interest in objects contained in a specific topic, which typically comprises a bulk of closely related objects. In this paper, we propose a local similarity method that has the ability to exploit multiple correlation structures between users who express their preferences for objects that are likely to have similar properties. For this, we use a clustering method to find groups of similar objects. Then we create a user-based similarity model for each cluster, which we named Cluster-based Local Similarity (CBLS) model. Each similarity model relies on rating normalization and resource allocation techniques that are sensitive to the ratings assigned to objects contained in the cluster. We performed experiments using two clustering algorithms (affinity propagation and K-Means) and compared the results with other neighborhood-based collaborative filtering approaches. Our numerical results on three benchmark datasets (MovieLens 100k, MovieLens 1M, and Netflix) demonstrate that the proposed method is competitive and outperforms traditional and state-of-the-art collaborative filtering-based similarity models in terms of accuracy metrics like mean absolute error (MAE) and root-mean-square error (RMSE).