基于社会正则的行为定向

发布时间：2018-03-14 14:03

本文选题：行为定向　切入点：泊松回归　出处：《计算机研究与发展》2013年S2期 　论文类型：期刊论文

【摘要】：行为定向又称为兴趣定向,是在线广告的一种有效投放手段,主要指网络媒体或广告联盟网络根据用户的历史行为信息为用户投放其感兴趣的广告.用户的历史行为信息包括搜索历史、网页浏览历史、广告点击和交易历史等,研究表明这些行为数据具有稀疏性、实时性和海量的特点.如何针对这些特点设计更好的行为定向方法已成为研究界和工业界面临的主要难题.针对行为数据具有稀疏性这一特点,提出一种新的融合社会信息与用户历史行为信息的行为定向方法.该方法的基本思想是基于Homophily理论:朋友之间对某一事物的喜好或态度具有一定的相似性,单个用户对该事物的喜好可通过其朋友们对该事物的态度来判断.首先将行为定向看作一个优化问题—采用泊松回归建模用户对广告的浏览和点击次数,并通过使数据似然最大的方法构建目标函数;然后将社交网络的朋友关系数据抽象为社会正则项,并将该正则项作为优化问题的一部分,以建模朋友之间的相互影响.本文中,我们根据朋友们喜好的相似度差异提出两个社会正则项:均值正则项———朋友们的喜好相似度较高;个体正则项———朋友们的喜好相似度差异较大.将这两个社会正则项分别加到上述优化问题中,对应均值模型和个体模型.实验表明这两个模型的性能要优于传统的只使用用户历史行为信息的行为定向方法;尤其当用户的历史行为数据稀疏时,我们的模型在点击率提升度上比传统方法提高了5%的精度.
[Abstract]:Behavioral orientation, also known as interest orientation, is an effective means of delivering online advertising. Mainly refers to the network media or the advertising alliance network according to the user's historical behavior information to place for the user to be interested the advertisement. The user's historical behavior information includes the search history, the webpage browsing history, the advertisement click and the transaction history and so on. Research shows that these behavioral data are sparse, How to design better behavior orientation methods based on these characteristics has become a major problem in research and industry. This paper proposes a new behavioral orientation method which combines social information with user's historical behavior information. The basic idea of this method is based on Homophily theory: there is a certain similarity between friends' preferences or attitudes towards a certain thing. A single user's preference for the thing can be judged by the attitude of his friends to it. First, behavioral orientation is considered as an optimization problem-Poisson regression is used to model users' views and clicks on advertisements. The objective function is constructed by the method of maximum likelihood of data, and then the friend relation data of social network is abstracted as a social regular term, and the canonical term is regarded as part of the optimization problem to model the interaction between friends. According to the similarity differences of friends' preferences, we propose two social regulars: the mean regularity--which is similar to friends' preferences; Individual canonical terms-- the similarity of friends' preferences is quite different. The two social canonical terms are added to the above optimization problem respectively. The experimental results show that the performance of the two models is superior to the traditional behavior orientation method which only uses the historical behavior information of the user, especially when the historical behavior data of the user is sparse. The accuracy of our model is 5% higher than that of the traditional method.
【作者单位】：中国科学院计算技术研究所;中国科学院信息工程研究所;
【基金】：国家“八六三”高技术研究发展计划基金项目(2011AA010703,2011AA01A103) 中国科学院战略性先导科技专项(XDA06030200) 国家自然科学基金项目(61003167)
【分类号】：TP391.3

【参考文献】