社交网络用户交互模型及行为偏好预测研究

发布时间：2018-03-31 13:44

本文选题：用户行为分析　切入点：偏好预测　出处：《北京邮电大学》2014年博士论文

【摘要】：互联网与移动通信技术的发展为各行各业带来了创新热潮,而随着以用户为中心理念的渗透,对用户行为的分析和预测已经成为提升用户体验的重要手段。进一步,各种社交网络的流行及智能终端的普及,为分析和预测用户的行为和偏好提供了海量的、真实的数据基础。然而,基于社交网络的用户行为分析和预测还存在一些问题需要解决。首先,在社交用户交互网络模型的构建方面,包括用户交互关系的生成和偏好一致性度量两个问题。一方面,社交网络中大部分用户的社交关系是残缺的、稀疏的,而这制约了用户行为预测的覆盖率和正确率。另一方面,用户基于相似的兴趣建立社交关系,然而,用户之间的兴趣存在差异。其次,社交影响作为社交网络中一个重要的特征被广泛的应用于预测模型中。在社交网络中,用户的行为、观念、看法、思想等,往往容易受到其社交关系的影响,而且这种社交影响还会沿着用户之间的关系链进行传播,如著名的“三度影响力”理论。如何衡量社交网络中用户之间的社交影响以及计算社交影响的传播是基于社交影响预测模型的关键。此外,如何从宏观上和微观上综合对用户行为偏好预测模型的预测结果进行有效评估也是一个有意义的研究问题。针对上述的用户交互网络构建、用户预测建模以及评估模型,论文围绕社交网络中的用户行为偏好预测主题展开研究。借助用户的图相似性提取用户的潜在关系,利用用户的兴趣数据设计有效的用户偏好一致性度量算法。在用户行为预测模型方面,对局部的、多样性的用户社交影响力计算模型进行了研究,同时,设计了一种可视化的、能够从微观层面评估预测算法性能的评估方法。本文的研究内容及主要贡献如下：由于社交网络的特性、用户的时间和精力等因素的限制,用户的社交关系往往是残缺的。对于大多数用户来说,其社交关系是非常稀疏的,导致了用户社交交互网络关系的残缺,直接在残缺的社交关系网络上进行预测会降低预测结果的覆盖率和正确率。基于节点相似度的方法是一种最简单且流行的用户关系挖掘方法,然而,其预测精度还有待于进一步提高。考虑到弱关系对于用户的链接概率的重要作用,论文提出一种基于网络节点中心性和弱关系理论的用户潜在关系挖掘算法,以提升用户缺失关系提取的准确性。社交网络中,用户兴趣是用户社交圈形成和维持的纽带,对用户偏好一致性的准确度量,可以提升用户行为预测的精度。而现有的用户偏好相似性计算方法存在准确性和区分度低等缺陷,不能很好的表征用户之间的偏好相似性,因此,在现有方法基础上,论文提出一种新颖的启发式用户相似性计算模型。该模型综合考虑影响用户偏好的微观因素和宏观因素,进一步提升了用户相似性的准确性,并且使得用户之间的相似性具有高度的可区分性。社交影响及其传播作为社交网络的一个重要特性,得到了研究者的广泛认可和研究兴趣。用户的行为、思想、决策等经常受到社交好友的影响,而且社交影响会沿着社交关系进行传播,通过对社交用户之间的社交影响及其传播的把握,可以帮助分析和预测用户的行为趋势。在社交影响计算方面,现有方法要么是缺乏多样性的全局社交影响,要么需要知道网络的整体信息。为此,论文分别提出一种基于局部节点网络拓扑和局部用户交互的社交影响计算方法,该方法将社交影响的计算限制在单个用户的邻居范围内,提取的社交影响是局部的、多样的,而且计算量小,同时采用最短及最大传播路径策略来建模社交影响的传播。有效的评估模型可以帮助在不同场景下选择最合适的预测算法,而大部分评估准则仅仅给出一个综合评估结果,并且建立在评估结果服从钟形分布的基础上,存在评估粒度不够细、不能随用户体验粒度进行调整等不足。本文研究发现很多预测结果的分布不符合钟形分布,而是近似服从一种幂率分布。因此,论文提出了一种累积概率分布模型的可视化评估方法,可以从更加细粒度的层面对不同的预测算法进行对比。同时基于该累积概率分布计算评估期望值,这样能够根据用户体验的粒度对预测结果进行离散化处理,因此计算得到的评估期望更加符合用户真实的体验和场景需求。
[Abstract]:The development of Internet and mobile communication technology have brought innovation boom for all walks of life, and with the development of user centered concept, analysis and prediction of user behavior has become an important means to enhance the user experience. Further, the popularity of social networks and intelligent terminals, providing analysis and predict the behavior of users and preference for massive, real data based.
However, user behavior analysis and prediction based on social network there are still some problems to be solved. First of all, in the construction of social network user interaction model, including measurement of two problems of user interaction between generation and preference consistency. On the one hand, most of the social network of users of social relations is incomplete, sparse but, this restricts the coverage of user behavior prediction and correct rate. On the other hand, similar user interests based on social relationship, however, there are differences between the user interest. Secondly, social influence as an important feature of the social network is widely used in the prediction model. In social networks, user behavior, ideas, opinions, ideas and so on, are easy to be affected by the social relations, and the social impact will be along the chain relationship between users of communication, such as the famous "three degrees. The force "theory. How to measure the effect of social communication between users in a social network and the social influence is the key calculation prediction model based on social influence. In addition, how to predict from the macro and micro integration of user behavior preference prediction model results in effective evaluation is a significant research problem.
According to the construction of network user interaction, the user modeling and evaluation model, based on the preferences of the user behavior in social network prediction themes of study. Through the user's relationship to extract the user map of the potential similarity metric algorithm based on user preference data user interest to design effective consistency. Prediction model in terms of user behavior, to part of the diversity of users of social influence model is studied, and a visual design, a method to evaluate the algorithm performance evaluation from the micro level. The research content of this paper and the main contributions are as follows:
Due to the characteristics of social network users, such as time and energy constraints, the user's social relationship is often incomplete. For most users, the social relationship is very sparse, leading to users of social interaction network is incomplete, direct relations network forecast will reduce the coverage prediction results and correct the rate of incomplete agency. Method based on node similarity mining method, one of the most simple and popular user relationship however, the prediction accuracy remains to be further improved. Considering the important role of weak ties for the link probability of users, this paper proposes a algorithm for mining user potential relationship network centrality and weak relationship based on the theory, to enhance the user accuracy. The lack of relation extraction
Social network users, user interest is the social circle to form and maintain ties, to accurately measure the consistency of user preferences, user behavior can enhance the prediction accuracy and user preference. The existing similarity calculation method in accuracy and discrimination of low defect, the user can not be a good characterization between the preference similarity, therefore, based on the existing method, this paper proposes a similarity calculation model and a heuristic novel. The model considering the micro factors and macro factors affecting user preferences, to further improve the accuracy of user similarity, and the similarity between users is highly distinguishable.
Social influence and communication as an important characteristic of the social network, has been widely recognized by researchers and research interest. The user's behavior, thinking, decision-making is often influenced by social friends, and social influence will spread along the social relations, the social impact of social and communication between users to grasp the behavior trend analysis and forecasting can help users. In the calculation of social effects, existing methods either lack of global social effects of diversity, or need to know the whole information network. Therefore, put forward a calculation method of the effect of local node network topology and user interaction based on the local social respectively, the method of calculating social influence Limited to a single user's neighbors within the scope of social impact extraction is local, diversity, and a small amount of calculation, the shortest and the maximum propagation The path strategy is used to model the communication of social impact.
The model can help to choose the most suitable prediction algorithm in different scenarios and the effective evaluation and evaluation criteria given only a most comprehensive evaluation results, and based on evaluation results based on the distribution of the bell to assess the size, fine enough, not enough with the user experience granularity adjustment. This study found that many predicted distribution the results do not conform to the bell shaped distribution, but approximately obey a power-law distribution. Therefore, this paper proposes a visual assessment method of cumulative probability distribution model, can be compared with different prediction algorithms from a more fine-grained layer. At the same time the evaluation of expected value of cumulative probability distribution based on this, according to the size of the user experience the forecast results are discretized, therefore the calculated expected assessment more in line with the user experience and the real needs of the scene.

【学位授予单位】：北京邮电大学
【学位级别】：博士
【学位授予年份】：2014
【分类号】：TP393.09

【参考文献】