【作者】 张伟；

【导师】 王建勇；

【作者基本信息】 清华大学 ， 计算机科学与技术， 2016， 博士

【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。更多还原

【Abstract】 Location-based social medias began to emerge with the rapid development of mobile Internet technology and widespread popularity of smart mobile devices,and have produced a large amount of location-aware user behavior data.In the era of big data,personalized recommendation,which models user behavior data,is the core technology for helping users to overcome the information overload problem and promoting the development of enterprise precision marketing.Traditional recommendation methods do not fully consider location information and its interactions with a variety of heterogeneous data,thus they can not obtain good recommendation results in this type of social media.This thesis aims at conducting systematic research on four typical user modeling and recommendation problems in location-based social medias based on the theories of data mining and machine learning to promote the further study of location-based social media in research field and deployment of recommendation applications in industry field.The research questions and technical contributions of this thesis are summarized as follows,1.Event-based group recommendation: to solve the problem that traditional group recommendation methods are proposed only for online virtual groups,this thesis fully explores the influence of location factors on users’ participation in groups and proposes a unified model,i.e.,PTARMIGAN,which combines the designed location-based features standard recommendation methods.The experimental results show that the proposed model is better than several other related methods in group recommendation,and further verify that considering location information is beneficial for predicting users to join event-based groups.2.Cold-start local event recommendation: to overcome the shortages that previous studies for event recommendation neglecting to address the cold-start events,this thesis proposes a collective Bayesian Poisson factorization model,i.e.,CBPF.This model fully considers heterogeneous information such as organizers of events,introduction of events,and venues of events,and designs an effective learning algorithm to obtain representations of cold-start events.The experimental results demonstrate CBPF outperforms several previous methods,and find organizers of events have the largest impact on whether users to join events.3.Time-aware next location recommendation: to exploit the joint influence of tem-poral factors,social relations,and current locations for user’s preference to next locations,this thesis proposes LTSCR by incorporating temporal information and social relation into the basic collaborative retrieval model.It regards a target user,the location where he currently stays,and temporal information as an implicit query,and ranking locations according to the relevance scores between the user and locations.The experimental results show that incorporating spatial-temporal information and social relations into collaborative retrieval model can obtain better performance on this task.4.User-item connected review modeling and rating prediction: To capture the characteristic that topics of a review are associated with its belonging user and location in location and review-based social medias,this thesis develops prior-based dual additive latent Dirichlet allocation(PDA-LDA)model.It associates parameters of Dirichlet distribution with user and item topic factors to influence topic distribution of reviews.The experimental results show PDA-LDA behaves well for text modeling and its generated topic features can benefit rating prediction.更多还原