Group recommender system based on genre preference focusing on reducing the clustering cost

초록

The most significant advantage of the group recommender system over personalization is the low computational cost because the former analyzes the preferences of many users at once by integrating their preferences. The clustering step is the most time-consuming part of the entire process in a group recommender system. Existing studies either measured the similarities among all users or utilized a clustering algorithm based on the item preference vector to form the groups. However, these existing clustering methods overlooked the clustering cost, and the time complexity was not significantly better than that for personalized recommendations. Therefore, we propose a group recommender system based on the genre preferences of users to dramatically reduce the clustering cost. First, we define a genre preference vector and cluster the groups using this vector. Our group recommender system can reduce the time complexity more efficiently because the number of genres is significantly smaller than the number of items. In addition, we propose a new item preference along with genre weight to subdivide the preferences of users. The evaluation results show that the genre-based group recommender system significantly improves the time efficiency in terms of clustering. Clustering time was about five times faster when using k-means. In addition, for the Gaussian mixture model (GMM), it was about fifty times faster in MovieLens 100 k and about five hundred times faster in Last.fm. The normalized discounted cumulative gain (NDCG) (i.e., accuracy) is not much different from that of the item-based existing studies and is even higher when the number of users is low in a group in MovieLens 100 k.

키워드