计算机科学与探索 ›› 2017, Vol. 11 ›› Issue (3): 373-381.DOI: 10.3778/j.issn.1673-9418.1603054

• 学术研究 • 上一篇    下一篇

融合项目标签信息面向排序的社会化推荐算法

练绪宝1,林鸿飞1+,徐  博1,林  原2   

  1. 1. 大连理工大学 计算机科学与技术学院,辽宁 大连 116024
    2. 大连理工大学 公共管理与法学学院,辽宁 大连 116024
  • 出版日期:2017-03-01 发布日期:2017-03-09

Rank-Oriented Social Recommendation Algorithm with Item Tag Information

LIAN Xubao1, LIN Hongfei1+, XU Bo1, LIN Yuan2   

  1. 1. School of Computer Science and Technology, Dalian University of Technology, Dalian, Liaoning 116024, China
    2. School of Public Administration and Law, Dalian University of Technology, Dalian, Liaoning 116024, China
  • Online:2017-03-01 Published:2017-03-09

摘要: 近年来,推荐系统越来越受到人们的关注,按照应用场景主要分为评分预测和Top-K推荐。考虑到传统评分推荐系统和Top-K排序推荐系统只考虑用户和项目的二元评分信息,具有一定的局限性,因此扩展了一种基于列表排序学习的矩阵分解方法。一方面,充分考虑用户之间关注关系。首先通过用户之间的关注关系计算用户之间的信任度,接着通过用户之间的信任度在原始模型的损失函数中添加用户社交约束项,使相互信任的用户偏好向量尽可能接近。另一方面,计算项目所拥有标签的权重,并以此计算项目之间的标签相似度,再将项目的标签约束项添加至损失函数中。在真实Epinions和百度电影数据集中的实验结果表明,该方法的NDCG值和原始模型相比具有一定的提高,有效地提高了推荐准确率。

关键词: 推荐系统, 社交网络, 标签系统, 排序学习, 矩阵分解

Abstract: In recent years, recommender system has attracted more and more attention. According to application scenario, recommender system can be divided into rating prediction and Top-K recommendation. Since traditional rating prediction and Top-K recommendation only consider limited dual rating information between users and items, this paper extends a list-wise learning to rank-based matrix factorization method. On one hand, the method fully considers the focusing relationship among users. At first, compute trust values between users based on users’ focusing relationship, then add trust matrix into the original loss function as a social penalty term to make users’ preference vectors as near as possible. On the other hand, the method computes the weights of tags of items, based on which to compute the tag similarities between items, and then add the item tag penalty term to the loss function for training the model. The experimental results on the real Epinions and BaiduMovie datasets show that the proposed method outperforms several traditional methods, especially on the NDCG value, improving the recommendation accuracy effectively.

Key words: recommender system, social networks, tag system, learning to rank, matrix factorization