2010 年 30 巻 4 号 p. 125-133
Collaborative filtering methods are frequently used for predicting users' preferences in recommender systems, such as those used for recommending movies, music, or articles. These methods have a large impact on businesses, because the volume of content they offer is tremendous, and it is important to support users' ability to make informed choices from among the available content. To increase sales and improve customer loyalty, many e-commerce companies, such as Amazon and Netflix, have adopted recommender systems. However, these companies generally rely on user ratings for the content they offer, and it is usually difficult or expensive to obtain such ratings data. Hence, we need a high-quality recommender system that uses only binary data, such as historical purchasing data, without ratings. Binary data, however, is very simple, and it is therefore difficult to express a relationship in detail between users and items by using only such simple methods. This paper proposes a recommender system based on a graph-partitioning method to solve the problems through a two-phase approach: We generate a model that expresses the relationships between various items and implements an appropriate grouping by using a graph-partitioning method. We then use our proposed algorithms to determine accurate recommendations. A comparison of our results with those obtained from traditional methods reveals that our method is more practical for businesses usage.