Journal of Frontiers of Computer Science and Technology ›› 2017, Vol. 11 ›› Issue (6): 875-886.DOI: 10.3778/j.issn.1673-9418.1605045

Previous Articles     Next Articles

Comparison Study of Collaborative Filtering Algorithms Based on Quadripartite Graph

MOU Binhao1, ZHANG Zhiheng2, ZHANG Lin1, MIN Fan1+   

  1. 1. School of Computer Science, Southwest Petroleum University, Chengdu 610500, China
    2. School of Sciences, Southwest Petroleum University, Chengdu 610500, China
  • Online:2017-06-01 Published:2017-06-07

基于四部图的协同过滤推荐算法比较研究

牟斌皓1,张智恒2,张  林1,闵  帆1+   

  1. 1. 西南石油大学 计算机科学学院,成都 610500
    2. 西南石油大学 理学院,成都 610500

Abstract: A recommender system often collects information about user profiles, item attributes and explicit ratings of users to items, which are further used to make predictions about unknown ratings. This paper constructs a quadripartite graph about the information and acquires ten algorithms from different parts of the graph. The first two algorithms are the classical user- and item-based collaborative filtering and only take into account the rating information. Four more algorithms take user or item as center and use relevant tags to compute user or item similarity. To extend the previous four algorithms, four more algorithms take into account the user-item relationship along with tag information. This paper compares the time complexity of different algorithms on two MovieLens data sets, and uses MAE (mean absolute error) and RMSE (root-mean-square error) metrics to evaluate the performance of different    algorithms. The experimental results demonstrate that the similarity of items is more reliable than that of users, and item tags are more useful than user tags. Besides, some simple linear integrations of different information are capable of enhancing recommendation performance.

Key words: recommender system, collaborative filtering, quadripartite graph, collaborative filtering tag

摘要: 推荐系统通常利用商品属性、用户信息以及用户对商品的已有评分来获取用户或者商品之间的相似度,进而预测未知评分。构造了关于这些信息的四部图,然后根据图中不同部分的组合获得了10类推荐算法,并比较了它们的时间复杂度。前两类算法基于用户与商品之间的关系,为经典的协同过滤算法。中间4类算法以用户或商品为中心,利用相应的标签信息进行相似度的计算并预测评分。后4类算法为中间4类算法的部分拓展,进一步考虑了评分信息。以MAE(mean absolute error)和RMSE(root-mean-square error)为评价指标,在两个MovieLens数据集上的测试结果表明,商品之间的相似度比用户之间的相似度更可靠,商品标签也比用户标签更有用,而且某些信息的简单线性组合可以提高推荐质量。

关键词: 推荐系统, 协同过滤, 四部图, 协同过滤标签