Abstract: Finding close relationships among users in social networks for recommendation services is one of hot research issues about social behavior relationships in both academic and industry domain. So far, most researches of behavior relationships in social networks are limited only to “follow” relationship, which results in the discovery of user relationships is inaccurate and incomplete. Fortunately, a special type of relationship supported by Twitter, i.e., “@(mention)” relationship, has been paid lots of attention for exploring latent reasonable behavior relationships due to its characteristics of revealing more closely relationships than the “follow” relationship. In this light, this paper proposes a new concept, UPBR (user pair behavior relationship). Two users producing similar semantic behaviors such as food, travel and shopping in some approximate geography space will have a UPBR relationship. Further, this paper also proposes an inferring model named UPBR-IM correspondingly for verifying UPBR. UPBR-IM infers not only activity similarity of semantic behaviors depending on “@” Tweets text, but also location similarity by means of maximum likelihood estimation method. Finally, this paper uses both similarities of activity and location to justify whether a pair of users have a UPBR relationship. The extensive experiment results validate the feasibility and effectiveness of the UPBR-IM model.

Key words: behavior relationship, semantic activity, similarity computing, geography position, maximum likelihood estimation, relationship inference

摘要： 如何发现具有紧密关系的用户并为其提供信息推荐服务，是目前学术界和工业界关于社交网用户行为关系研究的热点问题之一。迄今为止，大部分社交网用户行为关系研究主要局限于用户间“关注（follow）”行为上，导致用户间关系的发现尚不够准确和完善。在社交网应用之一的Twitter平台中，“@（mention）”关系相比于一般“关注”关系能更准确反映用户间紧密关系程度，因此提出了一种新的用户间相似关系，即UPBR（user pair behavior relationship）关系，该关系体现了在相近地理位置进行相似活动的语义行为，如餐饮、旅行、购物等；根据该关系提出了一种UPBR推演模型，即UPBR-IM，该模型一方面通过用户上传的@Tweet文本来推演用户语义行为活动相似性，另一方面则利用最大似然估计对文本发布位置进行概率最大化计算来推演用户物理位置相似性；最后结合二者结果判断用户间是否存在UPBR关系，从而实现高质量的信息推荐服务。扩展性的实验结果验证了该模型是可行和有效的。

关键词: 行为关系, 语义活动, 相似性计算, 物理位置, 最大似然估计, 关系推演