基于位置社交网络的用户行为建模与研究

发布时间：2018-01-06 04:15

本文关键词：基于位置社交网络的用户行为建模与研究　出处：《中国科学技术大学》2017年硕士论文　论文类型：学位论文

【摘要】：近年来,随着移动互联网的快速扩展和定位技术的日趋成熟,与位置社交网络相关的服务平台和信息被广泛应用于生活中。位置服务的广泛应用使得大量的位置数据得以积淀下来,这为挖掘位置数据背后用户的行为偏好提供了有力的支撑。通过分析用户的行为偏好,所构建的位置社交平台可以更好地便利人们的生活与出行,同时有关于用户偏好的分析结果也可以给予商家和相关行业的决策者更有益的建议和指导。因此,本文的工作重点是从现在和未来两个角度出发,挖掘和分析用户的行为偏好,从而进行兴趣点推荐和位置预测。虽然位置社交网络提供了丰富的位置数据来源,但是位置数据本身的异构性和稀疏性等特点给现有的推荐和预测方法带来了诸多挑战。针对位置数据的这一系列特点和存在的挑战,本文分别提出了相应的方法来更好地应对在推荐和预测问题建模过程中遇到的相关情况。具体来说包含以下两个方面:1.针对兴趣点推荐问题,本文构建了一个基于多源异构信息的混合兴趣点推荐模型。位置社交网络中蕴含着丰富的实体和关联关系,体现在位置数据上就是丰富的多源异构信息。通过合理的建模和算法设计来有效地整合这些信息可以改善兴趣点推荐的实际效果。针对位置社交网络中的多源异构信息,本文提出了一种基于用户虚拟兴趣和现实距离相结合的混合兴趣点推荐方法。具体来说,本文采用核密度估计的方法对地理空间距离来进行度量,使用基于好友和有共同签到地点的用户的协同过滤方法来衡量好友和兴趣相似的其他用户对于用户本身对兴趣点的心理认同度的影响,同时使用基于用户和兴趣点文本聚集的概率话题模型来挖掘用户和兴趣点的偏好,从而对用户虚拟兴趣中可解释的部分进行建模。相应的,本文使用概率隐因子模型对用户虚拟兴趣中不可解释的部分加以建模。最终本文将上述子模块有机地结合起来得到混合兴趣点推荐模型。本文在两个典型的位置数据集上进行了充分的实验,实验结果表明本文提出的混合兴趣点推荐算法优于当前已有的兴趣点推荐算法。此外,模型还具有更准确的预测性和很好的健壮性等优势。2.针对地点预测问题,本文提出了一种基于签到序列的隐话题向量位置预测模型。研究表明,位置社交网络中用户的行为偏好具有很强的规律性和可预测性,并且和用户与地点所在的情境密切相关。对于大多数用户来说,其签到记录相比于整个数据的分布而言具有很强的稀疏性。因此如何针对位置数据的上述特点构建预测模型来进行地点预测是一个亟待解决的重要问题。本文提出了一种基于签到序列的隐话题模型。具体来说,对于位置社交网络中的地理空间信息,本文采用基于区域的高斯分布模型进行建模。为了缓解社交关系稀疏性对预测结果的影响,本文对用户的社交关系进行了扩展。同时本文把基于上下文的词向量模型和基于时间的主题模型结合起来,构建隐话题向量模型来对用户签到行为的情境进行建模。对于其签到的规律性行为,本文对连续时间进行了横向与纵向的分割,把连续时间离散化。综合上述建模方法可以得到用户在不同时间模式下的兴趣偏好表示以及地点的表征向量,从而有效地预测下一时间模式下用户访问的地点。本文在典型的位置数据集上的实验结果表明与传统的地点预测方法相比,本文提出的模型具有更高的准确性。
[Abstract]:In recent years, with the rapid expansion of the Internet and mobile positioning technology is becoming more and more mature, and the position related social network service platform and information is widely used in daily life. Widely used location service that the position data settled and provides a strong support for the behavior preference mining. Through the data behind the user location analysis of user preferences, social position of the platform can better facilitate people's life and travel at the same time, there are more useful advice and guidance on user preference results can also be given to related businesses and industry decision makers. Therefore, the emphasis of this paper is to start from now and in the future two aspects of mining and analysis of user behavior, so as to predict the point of interest and recommended position. Although the location of social network provides a rich source of data location, but The location of the data itself characteristics of heterogeneous and sparseness and recommend the existing prediction methods have brought many challenges. For this series of characteristics of the position data and the challenges, this paper puts forward the corresponding methods to better respond to the relevant circumstances encountered in the process of the prediction and recommendation problem specifically includes modeling. The following two aspects: 1. to the point of interest problems is recommended, this paper constructs a hybrid recommendation model of multi-source heterogeneous information based on social network position. Points of interest are rich in entity and relationship, reflected in the position of the data is rich in multi-source heterogeneous information. Through the modeling and design of reasonable algorithm to effectively integrate this information can improve the actual effect of interest recommendation. For multi-source heterogeneous information position in social networks, this paper proposes a method based on user interest and virtual The real distance of combining interest recommendation methods. Specifically, this paper uses the method of kernel density estimation for spatial distance measurement, based on the use of friends and common collaborative filtering method to measure the user sign in place of friends and other users with similar interests to the user itself to the point of interest is the influence of psychological identity at the same time, the use of probabilistic topic model and user interest aggregation to text mining user preferences and points of interest, in order to model the interpretation of part of the user interest. The corresponding virtual modeling of virtual users, it can not explain the interest in this part of the use of probabilistic latent factor model. Finally the sub module organically mixed interest recommendation model. This paper has carried on the experiment in two typical position data sets, experimental results table The proposed hybrid algorithm is better than the current recommended interest interest recommendation algorithm. In addition, the model has more accurate prediction and good robustness for the.2. advantage location prediction problem, this paper proposes a prediction model based on hidden topic vector position sign sequence. The results show that the position of the user in a social network behavior preference has strong regularity and predictability, and closely related to the user's location and context. For most users, the attendance record distribution compared to the entire data sparsity has very strong. So how to according to the characteristics of the position data prediction model is constructed to place forecasting is an important problem to be solved. This paper proposes a topic model based on implicit sign sequence. Specifically, the position of social networks in the geographical space The information modeling of Gauss distribution model based on region. In order to alleviate the impact of social relations the sparsity of the forecast results, the relationships of the users of the expansion. At the same time the word context vector model and topic model based on time based on user behavior, to construct the context modeling sign the hidden topic vector model. For the regularity of the sign of sexual behavior, the horizontal and vertical segmentation of continuous time, continuous time discretization. The modeling method can get the user preference vector characterization in different time under the mode of representation and location, which can effectively predict the next time the user mode to access a location. Based on the position data of the typical set of experimental results show that compared with the traditional location prediction method, the model proposed in this paper. There is a higher accuracy.

【学位授予单位】：中国科学技术大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.3

【参考文献】