• 中国计算机学会会刊
  • 中国科技核心期刊
  • 中文核心期刊

计算机工程与科学

• 论文 • 上一篇    下一篇

幂律特性在新浪微博个性化推荐中的应用研究

罗斌,陈翔   

  1. (北京理工大学管理与经济学院,北京 100081)
  • 收稿日期:2016-05-18 修回日期:2016-09-22 出版日期:2018-04-25 发布日期:2018-04-25
  • 基金资助:

    国家自然科学基金(71102111)

Application of power-law characteristics in
Sina-weibo personalized recommendation

LUO Bin,CHEN Xiang   

  1. (School of Management and Economics,Beijing Institute of Technology,Beijing 100081,China)
  • Received:2016-05-18 Revised:2016-09-22 Online:2018-04-25 Published:2018-04-25

摘要:

在社交网络数据中,幂率分布是数据的基本规律,幂率分布的长尾部分数据有明显的稀疏性,长尾推荐一直是推荐系统的挑战,而冷启动、数据稀疏和覆盖率这些问题也是推荐系统的重要研究内容。通过分析数据幂律分布的特性,在研究社交网络个性化推荐方法的基础上,结合社交网络用户行为数据反映出来的幂律分布特性,通过极大似然估计数据幂律分布的标度值。结合幂率特性改进了相似度计算方法,提出了一种基于幂率特性的混合推荐方法PowerLawCF。最后,使用新浪微博的用户签到数据进行对比分析,PowerLawCF算法的推荐效果有显著提升,提高了长尾推荐的效果,对推荐系统的数据稀疏性和冷启动问题解决效果较好,说明基于幂律分布特征的推荐方法在推荐系统中的应用有积极的效果。

关键词: 幂律分布, 长尾分布, 社交网络, 协同过滤推荐, 数据稀疏

Abstract:

Powerlaw distribution is one of the basic laws of social network data. The long tail part of the data with powerlaw distribution has obvious sparsity. Long tail is always a challenge of recommendation systems. The coldstart, sparsity and coverage problems of recommended systems are important research contents. This paper analyzes the powerlaw characteristics of data, based on studying the personalized recommendation methods of social networks, combining the powerlaw distribution characteristics derived from the behaviors of social network users, the maximum likelihood estimation is used to calculate the scale values of powerlaw distribution. The similarity calculation method is improved by combining the powerlaw characteristics, and a hybrid recommended method based on powerlaw, named PowerLawCF (Collaboration Filter) is proposed. Finally, we use Weibo data to make comparative analysis. PowerLawCF enhances the recommended effect significantly, improves the effect of long tail recommendation, and better solves the cold-start, sparsity and coverage problems of recommended systems, which shows that . The proposed method based on powerlaw distribution characteristics plays a better role in recommendation systems.
 

Key words: power-law distribution, long tail distribution, social networks, collaborative filtering recommendation, sparsity