基于社会化媒体的若干兴趣点推荐关键技术研究
发布时间：2019-01-28 13:04
【摘要】：随着以Web2.0技术为基础的社会化媒体的兴起,基于位置的社交网络(LBSN,Location Based Social Network)服务、各种移动端社会化媒体的出现以及城市的快速发展,兴趣点(POI,Point-of-Interest)的数量也随之增长,人们通常喜欢探索城市与邻近的地方,根据自已的个人兴趣选择与自已偏好相关的兴趣点。基于位置的社交网络为研究人们移动行为提供了前所未有的机会,用户喜欢在这些基于位置的社交网络平台上,分享他们对各个地方的签到记录与兴趣爱好,以及他们对服务、产品的评价与体验,并且建立与维护他们的社会关系,从而展现自已的偏好与个性。这些基于位置的社交网络的创建者也更加重视对用户基础数据和行为数据进行采集、挖掘与分析,更好地理解用户的移动行为,从而更加了解他们的用户,利用兴趣点推荐改善用户体验并满足用户需求。同时社会化媒体的兴趣点推荐会面临一些新的问题:如何综合利用社会化媒体中的多样数据?如何解决用户签到数据的稀疏性?如何处理隐式的用户反馈与复杂的用户关系?如何应对用户生成内容的时效性?针对这些挑战,本文提出并设计一系列融合上下文信息的兴趣点推荐算法,提高并改善社会化媒体中的兴趣点推荐效果以及用户体验。本文创新工作如下:1.基于位置社交网络的上下文感知的兴趣点推荐。基于位置社交网络中的兴趣点签到矩阵是高稀疏的,用户兴趣随着不同时间与地理位置是动态变化的。针对此问题,本文提出一种上下文感知的概率矩阵分解兴趣点推荐算法。首先利用潜在狄利克雷分配(LDA,Latent Dirichlet Allocation)模型挖掘兴趣点相关的文本信息学习用户的兴趣话题生成兴趣相关分数;其次提出一种自适应带宽核评估方法构建地理相关性生成地理相关分数;然后通过用户社会关系的幂律分布构建社会相关性生成社会相关分数;结合用户的分类偏好与兴趣点的流行度构建分类相关性生成分类相关分数;将这四种相关分数进行分数匹配生成偏好分数;最后将其有效融合到概率矩阵分解模型(PMF, Probabilistic MatrixFactorization),生成用户感兴趣的兴趣点推荐列表。实验结果表明,该模型明显优于先进的NCPD算法,在Foursquare数据集上,准确率和召回率分别提高了 27%和24%;在Twitter数据集上,准确率和召回率分别提高了 26%和25%,显著提高了兴趣点推荐的精确度。2.基于用户签到行为的兴趣点推荐。目前缺乏一种综合分析地理影响、时间效应、社会相关性、内容信息和流行度影响这些因素共同作用的方法来处理兴趣点推荐稀疏性问题,特别是异地推荐场景。针对此问题,本文提出一种联合概率生成模型,第一个同时将上述因素进行有效融合的联合效应模型,模拟用户签到行为的决策过程,利用地理相关性设计一个良好的空间索引结构即空间金字塔,对当地偏好进行平滑优化,进一步缓解数据稀疏问题。该模型包括离线模型和在线推荐两个部分,支持本地和异地两种推荐场景,并利用一个可扩展的查询过程技术阈值算法加速在线推荐过程。实验结果表明该模型明显优于先进的SVDFeature算法,异地推荐场景中,在Foursquare数据集上,准确率和召回率分别提高了 24%和26%,在Twitter数据集上,准确率和召回率分别提高了 21%和23%,在豆瓣数据集上,准确率和召回率分别提高了 22%和24%;本地推荐场景中,在Foursquare数据集上,准确率和召回率分别提高了 14%和16%,在Twitter数据集上,准确率和召回率分别提高了 23%和20%,在豆瓣数据集上,准确率和召回率分别提高了 15%和17%,显著提高了兴趣点推荐的精确度。3.基于社会化媒体挖掘与可视化的兴趣点推荐。社会化媒体的社交网络中,图像还没有很好地被利用到兴趣点推荐研究。针对此问题,本文提出一种社会化媒体主题模型,充分利用Twitter的文本、图像、位置、时间和哈希标签这五个特征之间的内在关联性构建一个联合概率生成模型。并研究Twitter上的图像对兴趣点推荐的影响,解决噪声图像问题,预先定义三个标准:可视化一致性、可视化相关性与可视化多样性,利用卷积神经网络(CNN, Convolutional Neural Network)选择代表性的图像对兴趣点进行可视化。实验结果表明,该模型明显优于先进的TRM算法,在Twitter数据集上,平均准确率提高了 22%,显著提高了兴趣点推荐的精确度。
[Abstract]:With the rise of the social media based on the Web 2.0 technology, the location-based social network (LBSN, Location Based Social Network) service, the emergence of various mobile-end social media and the rapid development of the city, the number of points of interest (POI, Point-of-Interest) has also increased, People often like to explore cities and nearby places, and choose the points of interest related to their own preferences based on their personal interests. The location-based social network provides unprecedented opportunities for the study of people's mobile behavior, and users like to share their location-based social networking platforms, share their sign-in records and interests at various places, and their evaluation and experience of services, products, and establish and maintain their social relations so as to show their own preferences and personalities. The creators of these location-based social networks also pay more attention to the collection, mining, and analysis of user-based data and behavior data, to better understand the user's mobile behavior, to better understand their users, to use the point of interest to recommend to improve the user experience and to meet the user's needs. At the same time, the point of interest of the social media is recommended to face some new problems: how to make comprehensive use of the diverse data in the social media? How to solve the sparsity of user sign-in data? How to handle implicit user feedback and complex user relations? How to deal with the timeliness of user-generated content? In view of these challenges, a series of point-of-interest recommendation algorithms for integrating context information are proposed and designed, and the recommendation effect and user experience of the interest point in the social media are improved and improved. The innovative work of this paper is as follows: 1. A location-based social network-based context-aware point of interest recommendation. The sign-in matrix of the point of interest in the location-based social network is highly sparse, and the user's interest is dynamically changed with the different time and the geographic location. In this paper, a context-aware probability matrix decomposition point-of-interest recommendation algorithm is presented in this paper. firstly, using a potential Dirichlet allocation (LDA) model to mine the interest point-related text information to study the interest topic of the user to generate an interest-related score; secondly, a self-adaptive bandwidth core evaluation method is proposed to construct a geographic correlation to generate a geographic correlation score; then building a social correlation to generate a social correlation score through the power law distribution of the social relation of the user; building a classification correlation score according to the popularity of the user's classification preference and the popularity of the interest point; and carrying out score matching on the four related scores to generate a preference score; Finally, it is effectively fused to the probability matrix decomposition model (PMF, Probabble MatrixFactorization) to generate a list of the recommended points of interest for the user. The experimental results show that the model is better than the advanced NCPD algorithm, and the accuracy and recall rate are increased by 27% and 24% respectively on the Foursquare data set, and the accuracy and recall rate are increased by 26% and 25% respectively on the Twitter data set, and the recommended accuracy of the point of interest is significantly improved. Recommendation for the point of interest based on the user's sign-in behavior. At present, there is a lack of a method to comprehensively analyze the geographical influence, time effect, social relevance, content information and popularity influence the common function of these factors to treat the point of interest and to recommend the sparsity problem, in particular to the recommendation scene in different places. In order to solve this problem, a joint probability generation model is proposed, the first is the combined effect model of the effective fusion of the above factors, the decision-making process of the user sign-in behavior is simulated, and a good spatial index structure, i.e. the spatial pyramid, is designed by using the geographic relevance. The local preference is optimized to further alleviate the data sparse problem. The model includes the off-line model and the on-line recommendation two parts, supports both local and off-site recommendation scenarios, and uses an extensible query process technology threshold algorithm to accelerate the on-line recommendation process. The experimental results show that the model is superior to the advanced SVDFeature algorithm, and the accuracy and recall rate are increased by 24% and 26% respectively on the Foursquare data set, and the accuracy and recall rate are increased by 21% and 23% on the Twitter data set, respectively. The accuracy and recall rate increased by 22% and 24%, respectively. In the local recommendation scenario, the accuracy and recall rate increased by 14% and 16% respectively on the Foursquare data set, and the accuracy and recall rate were increased by 23% and 20% on the Twitter data set, respectively. The accuracy and recall rate increased by 15% and 17%, respectively, and the recommended accuracy of the point of interest was significantly improved. The recommendation of the interest point based on the social media mining and visualization. In the social network of the social media, the image has not been well used to the point of interest recommendation study. In order to solve this problem, this paper proposes a social media subject model, and makes full use of the inherent relationship between the five features of the text, image, location, time and hash tag of Twitter to construct a joint probability generation model. In this paper, the effect of the image on the recommendation of the point of interest is studied, the problem of the noise image is solved, and the three criteria are defined in advance: the visual consistency, the visual correlation and the visual diversity, and the representative image is selected by the convolution neural network (CNN, Convolutional Neural Network) to visualize the points of interest. The experimental results show that the model is better than the advanced TRM algorithm. On the Twitter data set, the average accuracy of the model is increased by 22%, and the recommended accuracy of the point of interest is significantly improved.
【学位授予单位】：北京邮电大学
【学位级别】：博士
【学位授予年份】：2017
【分类号】：TP391.3
本文编号：2417093
[Abstract]:With the rise of the social media based on the Web 2.0 technology, the location-based social network (LBSN, Location Based Social Network) service, the emergence of various mobile-end social media and the rapid development of the city, the number of points of interest (POI, Point-of-Interest) has also increased, People often like to explore cities and nearby places, and choose the points of interest related to their own preferences based on their personal interests. The location-based social network provides unprecedented opportunities for the study of people's mobile behavior, and users like to share their location-based social networking platforms, share their sign-in records and interests at various places, and their evaluation and experience of services, products, and establish and maintain their social relations so as to show their own preferences and personalities. The creators of these location-based social networks also pay more attention to the collection, mining, and analysis of user-based data and behavior data, to better understand the user's mobile behavior, to better understand their users, to use the point of interest to recommend to improve the user experience and to meet the user's needs. At the same time, the point of interest of the social media is recommended to face some new problems: how to make comprehensive use of the diverse data in the social media? How to solve the sparsity of user sign-in data? How to handle implicit user feedback and complex user relations? How to deal with the timeliness of user-generated content? In view of these challenges, a series of point-of-interest recommendation algorithms for integrating context information are proposed and designed, and the recommendation effect and user experience of the interest point in the social media are improved and improved. The innovative work of this paper is as follows: 1. A location-based social network-based context-aware point of interest recommendation. The sign-in matrix of the point of interest in the location-based social network is highly sparse, and the user's interest is dynamically changed with the different time and the geographic location. In this paper, a context-aware probability matrix decomposition point-of-interest recommendation algorithm is presented in this paper. firstly, using a potential Dirichlet allocation (LDA) model to mine the interest point-related text information to study the interest topic of the user to generate an interest-related score; secondly, a self-adaptive bandwidth core evaluation method is proposed to construct a geographic correlation to generate a geographic correlation score; then building a social correlation to generate a social correlation score through the power law distribution of the social relation of the user; building a classification correlation score according to the popularity of the user's classification preference and the popularity of the interest point; and carrying out score matching on the four related scores to generate a preference score; Finally, it is effectively fused to the probability matrix decomposition model (PMF, Probabble MatrixFactorization) to generate a list of the recommended points of interest for the user. The experimental results show that the model is better than the advanced NCPD algorithm, and the accuracy and recall rate are increased by 27% and 24% respectively on the Foursquare data set, and the accuracy and recall rate are increased by 26% and 25% respectively on the Twitter data set, and the recommended accuracy of the point of interest is significantly improved. Recommendation for the point of interest based on the user's sign-in behavior. At present, there is a lack of a method to comprehensively analyze the geographical influence, time effect, social relevance, content information and popularity influence the common function of these factors to treat the point of interest and to recommend the sparsity problem, in particular to the recommendation scene in different places. In order to solve this problem, a joint probability generation model is proposed, the first is the combined effect model of the effective fusion of the above factors, the decision-making process of the user sign-in behavior is simulated, and a good spatial index structure, i.e. the spatial pyramid, is designed by using the geographic relevance. The local preference is optimized to further alleviate the data sparse problem. The model includes the off-line model and the on-line recommendation two parts, supports both local and off-site recommendation scenarios, and uses an extensible query process technology threshold algorithm to accelerate the on-line recommendation process. The experimental results show that the model is superior to the advanced SVDFeature algorithm, and the accuracy and recall rate are increased by 24% and 26% respectively on the Foursquare data set, and the accuracy and recall rate are increased by 21% and 23% on the Twitter data set, respectively. The accuracy and recall rate increased by 22% and 24%, respectively. In the local recommendation scenario, the accuracy and recall rate increased by 14% and 16% respectively on the Foursquare data set, and the accuracy and recall rate were increased by 23% and 20% on the Twitter data set, respectively. The accuracy and recall rate increased by 15% and 17%, respectively, and the recommended accuracy of the point of interest was significantly improved. The recommendation of the interest point based on the social media mining and visualization. In the social network of the social media, the image has not been well used to the point of interest recommendation study. In order to solve this problem, this paper proposes a social media subject model, and makes full use of the inherent relationship between the five features of the text, image, location, time and hash tag of Twitter to construct a joint probability generation model. In this paper, the effect of the image on the recommendation of the point of interest is studied, the problem of the noise image is solved, and the three criteria are defined in advance: the visual consistency, the visual correlation and the visual diversity, and the representative image is selected by the convolution neural network (CNN, Convolutional Neural Network) to visualize the points of interest. The experimental results show that the model is better than the advanced TRM algorithm. On the Twitter data set, the average accuracy of the model is increased by 22%, and the recommended accuracy of the point of interest is significantly improved.
【学位授予单位】：北京邮电大学
【学位级别】：博士
【学位授予年份】：2017
【分类号】：TP391.3
【参考文献】
相关博士学位论文 前1条
1 阴红志;社会化媒体中若干时空相关的推荐问题研究[D];北京大学;2014年
，本文编号：2417093
本文链接：https://www.wllwen.com/kejilunwen/ruanjiangongchenglunwen/2417093.html
