ISSN 1000-1239 CN 11-1777/TP

计算机研究与发展 ›› 2015, Vol. 52 ›› Issue (5): 1014-1021.doi: 10.7544/issn1000-1239.2015.20131551

• 信息处理 • 上一篇    下一篇

社区热点微博推荐研究

彭泽环1,孙乐1,2,韩先培1,2,陈波1   

  1. 1(中国科学院软件研究所基础软件国家工程研究中心 北京 100190); 2(计算机科学国家重点实验室(中国科学院软件研究所) 北京 100190) (pengzehuan@yahoo.cn)
  • 出版日期: 2015-05-01
  • 基金资助: 
    基金项目:国家自然科学基金项目(61433015,61272324);国家“八六三”高技术研究发展计划基金项目(2015AA015405);网络文化与数字传播北京市重点实验室开放课题(ICDD201204)

Community Hot Statuses Recommendation

Peng Zehuan1, Sun Le1,2, Han Xianpei1,2, Chen Bo1   

  1. 1(National Engineering Research Center of Fundamental Software, Institute of Software, Chinese Academy of Sciences, Beijing 100190); 2(State Key Laboratory of Computer Science(Institute of Software, Chinese Academy of Sciences), Beijing 100190)
  • Online: 2015-05-01

摘要: 分析并总结了影响用户对特定微博兴趣的若干因素,在此基础上基于潜在因素模型提出了1个融合显式特征和潜在特征的社区热点微博推荐算法(community micro-blog recommendation, CMR),并将其用于发现微博兴趣社区热点信息.算法在3个兴趣社区上进行了实验,结果表明:1)融合2种特征信息的微博推荐效果好于使用单一特征信息的推荐;2)CMR的推荐效果好于基于转发次数的对照实验(micro-blog repost rank based recommendation, MRR);3)通过分析各个算法所推荐的微博内容,发现CMR倾向于为用户推荐兴趣社区相关微博,而MRR倾向于为用户推荐公共热点微博.

关键词: 微博, 推荐, 社区, 潜在因素模型, 信息过载

Abstract: Micro-blog recommendation is an effective technique to resolve the information overload problem in micro-blog systems. In this paper, we summarize and model several key factors which affect a user's interest on a specific status, including implicit features, content features (i.e., content similarity, user tags, and user's favorites), social network features, and status features. Based on the above features, we propose a community hot status recommendation algorithm—CMR (community micro-blog recommendation), which combines both explicit features and implicit features for better recommendation. Specifically, we propose a learning method to rank based framework, which learns a user's interest model of status from his preference data, including his retweets, favorites, comments, etc. Then new statuses are scored and ranked using the learned interest model. In order to measure our method's performance, we conduct a series of experiments in three community data sets (including NLP, Photography and Basketball). Experimental results show that: 1)by combining both implicit features and explicit features, our method achieves better recommendation performance than that using a single type of features; 2) compared with the MRR (micro-blog repost rank based recommendation), CMR gets better recommendation performance; 3) MRR prefers recommending hot statuss in the whole micro-blog system, in contrast CMR usually recommends community-specific statuses.

Key words: micro-blog, recommendation, community, latent factor model, information overloading

中图分类号: