基于位置的社交网络链接预测系统研究

发布时间：2018-05-01 11:13

本文选题：链接预测 + LBSN　；参考：《北京交通大学》2015年硕士论文

【摘要】：基于位置的社交网络(location-based social network,LBSN)提供了用户的在线网络关系和签到(check-in)的空间时间等多重信息,连接了虚拟网络和现实生活,不仅丰富了人们的网络生活,也为数据挖掘和移动互联网等领域提供了新的研究方向。LBSN的好友关系预测作为其中的一方面,通常考察网络结构和空间位置维度的信息,预测特征分析较单一。此外,LBSN具有网络结构稀疏、预测空间大等特点,也对系统预测性能提出了挑战。本文针对以上问题,在传统的基于用户相似性的链接预测方法基础上,提出新的LBSN链接预测系统框架,Brightkite和Gowalla网络数据仿真结果表明该系统具有良好的预测效果。本文的主要工作和贡献包括： 1.采用典型的基于位置的社交网络Brightkite和Gowalla数据集,挖掘LBSN网络结构和用户签到行为特征。分析表明用户签到次数、地点数以及签到地点的访问人数、访问量均呈长尾分布,同时发现部分相对孤立的用户和地点,需要对数据进一步处理。 2.去除地点关系网络(the Co-location Network)中的孤立点,解决“僵尸”用户问题,避免采用朋友-地点关系(the Co-located Friends Network)网络造成保留的用户过少。 3.采用Louvain算法对网络进行社区划分,在减小预测空间和计算时间的同时提高链接预测平均准确率。 4.从网络结构和用户签到行为多方面挖掘用户相似性,提出两类基于签到时间和签到频率的新的LBSN链接预测测度,统计分析各指标与用户间建立链接的平均概率相关关系。 5.建立基于相似性的LBSN链接预测系统框架,进行非监督和监督式链接预测仿真,分析各指标预测性能。实验结果表明,相比传统的基于网络结构和签到地点预测特征,系统加入基于签到时间和签到频率的预测特征后,整体预测效果明显改善,预测准确率和F1值最高分别提升15.5%和7.4%。
[Abstract]:Location-based social network-based LBSNs provide users with multiple information such as online network relationships and check-in space and time, connecting virtual networks and real life, not only enriching people's network life. It also provides a new research direction for data mining and mobile Internet. LBSN's friend relationship prediction is one of them. As one of them, the information of network structure and spatial location dimension is usually investigated, and the prediction feature analysis is relatively simple. In addition, LBSN has the characteristics of sparse network structure and large prediction space, which also challenges the predictability of the system. Aiming at the above problems, based on the traditional link prediction method based on user similarity, a new LBSN link prediction system framework, Brightkite and Gowalla network data simulation results show that the system has a good prediction effect. The main work and contributions of this paper include: 1. The typical location-based social network Brightkite and Gowalla datasets are used to mine LBSN network structure and user check-in behavior. The analysis shows that the number of users' check-in, the number of site points, the number of visitors and the number of visitors are all long tail distribution. At the same time, it is found that some isolated users and locations need to be processed further. 2. Remove outliers from the Co-location Network, solve the "zombie" user problem, and avoid using the Co-located Friends Network network to create too few reserved users. 3. The Louvain algorithm is used to divide the community of the network, which reduces the prediction space and computation time, and improves the average accuracy of link prediction. 4. In this paper, two kinds of new LBSN link prediction measures based on check-in time and check-in frequency are proposed to mine user similarity from network structure and user check-in behavior, and the average probability correlation between each index and user is analyzed statistically. 5. The LBSN link prediction system framework based on similarity is established, and the unsupervised and supervised link prediction simulation is carried out, and the performance of each index is analyzed. The experimental results show that, compared with the traditional prediction features based on network structure and check-in location, the overall prediction effect is obviously improved by adding the prediction features based on check-in time and check-in frequency. The highest predictive accuracy and F1 value were increased by 15.5% and 7.4%, respectively.
【学位授予单位】：北京交通大学
【学位级别】：硕士
【学位授予年份】：2015
【分类号】：TP393.09

【参考文献】