Tag Genome Aware Collaborative Filtering Based on Item Clustering for Recommendations

Item-based collaborative filtering has become the most popular algorithm in recommender systems which exploits relationships among items to predict users’ preferences for items. However, this algorithm suffers from data sparsity and poor varieties of information sources. It only contains information from rating data, lacking the items’ meta information such as classifications and characteristics, which make similarities measurement among items imprecise, consequently reduce accuracy and interpretability of results. In this paper, we incorporate tag genome information into item-based collaborative filtering using item clustering techniques, which provides a more objective and comprehensive description of items. In particular, we utilize relevance between items and tags to calculate similarities in a more precise way and integrate the results with item-based collaborative filtering through an optimization process. An innovative improved stochastic gradient descent algorithm is adopted to accelerate the optimization process, which lowers down the time complexity without sacrificing the accuracy of the model. Our experimental results demonstrate the effectiveness of our proposed algorithm by gaining lower rating prediction error than currently popular item-based collaborative filtering and SVD latent factor model, and also outperforms them in a more practical top K recommender system.