基于地点的社交媒体中用户建模与内容推荐

发布时间：2019-04-22 20:20

【摘要】：随着移动互联网技术的迅猛发展和智能移动设备的广泛普及,基于地点的社交媒体应运而生,并产生了大量与地点相关的用户行为数据。个性化推荐针对用户行为数据建模,是大数据时代下帮助用户克服信息过载和实现企业精准营销的核心技术。传统的推荐算法并未充分考虑地理位置信息及其与多种类型数据之间的交互关系,因而不能在该新媒体中获得理想的推荐效果。本文以数据挖掘和机器学习相关理论为基础,对基于地点的社会化媒体中4种代表性的用户建模和内容推荐问题开展了系统性研究,能够促进学术界对基于地点的社交媒体进行深度研究,并推动工业界部署相关推荐应用。本文的研究问题和技术贡献总结如下:1.基于聚会的小组推荐:为解决传统小组推荐方法只针对线上虚拟小组的问题,本文充分探究了地理位置因素对用户参加小组的影响,并提出了一个统一模型PTARMIGAN将设计的地点相关特征与经典推荐方法融合。实验结果表明该模型在小组推荐效果上比相关方法更佳,并验证了考虑地理位置信息有助于预测用户加入基于聚会的小组。2.冷启动同城活动推荐:为克服前人活动推荐研究中忽视对冷启动活动处理的不足,本文提出了联合贝叶斯泊松分解模型(CBPF)。它充分考虑了线下活动组织者、活动内容介绍、活动举办地点等异构信息,并设计了有效的学习算法得到冷启动活动的表示。实验结果验证了CBPF优于多个前人的方法,并发现活动组织者对于用户是否参加该活动影响最大。3.时间感知的下一地点推荐:为利用时间因素、社会关系和当前地点对用户下一地点偏好的共同影响,本文在基本协同检索模型的基础上提出了融入时间和社会关系的新模型LTSCR,它将目标用户、其所处地点与对应时间作为隐式查询,并依据该查询对候选地点打分排序。实验结果说明融入时空信息和社会关系的协同检索模型在该任务上效果更好。4.用户地点关联的评论建模与评分预测:为捕捉基于地点的点评型社交媒体中评论主题同时与所属用户和地点关联这一特点,本文开发了基于先验的对偶可加潜在狄利克雷分配模型(PDA-LDA)。该模型将用户和地点的主题因子与狄利克雷分布参数关联,从而影响评论的主题产生。实验结果说明该模型在文本建模上表现优异,且产生的主题分布特征有益于评分预测。
[Abstract]:With the rapid development of mobile Internet technology and the widespread popularity of intelligent mobile devices, location-based social media emerged as the times require, and generated a large number of location-related user behavior data. Personalized recommendation for user behavior data modeling is the core technology to help users overcome information overload and achieve accurate marketing in big data era. The traditional recommendation algorithm does not fully consider the geographic location information and its interaction with many kinds of data, so it can not get the ideal recommendation effect in the new media. Based on the theory of data mining and machine learning, this paper systematically studies four representative user modeling and content recommendation problems in location-based social media. Facilitate in-depth academic research on location-based social media and facilitate the deployment of recommendation applications in industry. The research questions and technical contributions in this paper are summarized as follows: 1. Party-based group recommendations: in order to solve the problem of traditional group recommendations only for online virtual groups, this article fully explores the impact of geographic factors on user participation in the group. A unified model, PTARMIGAN, is proposed to integrate the location-related features of the design with the classical recommendation methods. The experimental results show that the model is better than the correlation method in terms of group recommendation, and it is verified that considering geographic location information can help predict users to join a party-based group. 2. Recommendation of cold start in city: in order to overcome the shortage of neglecting the treatment of cold start activity in previous study of activity recommendation, a joint Bayesian Poisson decomposition model (CBPF).) is proposed in this paper. It takes full account of heterogeneous information such as offline event organizer, activity content introduction, event location and so on, and designs an effective learning algorithm to get the representation of cold start activity. The experimental results show that CBPF is superior to many predecessors, and it is found that the organizer of the event has the greatest influence on the user's participation in the activity. 3. Time-aware next location recommendation: in order to make use of the time factor, social relations and current location on the user's preference for the next location, this paper proposes a new model, LTSCR, which integrates time and social relations on the basis of the basic collaborative retrieval model. It takes the target user, its location and corresponding time as implicit query, and sorts the candidate location according to the query. The experimental results show that the cooperative retrieval model, which integrates spatio-temporal information and social relations, is more effective in this task. 4. Comment modeling and scoring prediction for user location association: to capture the feature that comment topics in location-based peer-reviewed social media are associated with their users and locations at the same time. A priori-based dual additive latent Dirichlet assignment model (PDA-LDA) is developed in this paper. The model correlates user and location theme factors with Dirichlet distribution parameters, thus affecting the topic generation of comments. The experimental results show that the model performs well in text modeling, and the topic distribution features are useful for scoring prediction.
【学位授予单位】：清华大学
【学位级别】：博士
【学位授予年份】：2016
【分类号】：TP391.3

【相似文献】