摘要:作为一种新兴的社交媒体用户交互服务，提及机制（mention mechanism）正在用户在线交互和网络信息传播方面扮演着重要角色.对用户提及行为的研究能够揭示用户的隐式偏好与其显式行为之间的联系，为信息传播监控、商业智能、个性化推荐等应用提供新的数据支撑.当前，对用户提及机制的探索多集中在其信息传播属性上，缺少从普通用户角度对其用户交互属性的学习.通过对普通用户提及行为的分析和建模构建一个推荐系统，为给定的社交媒体消息生成目标用户推荐.通过对大型真实社交媒体数据集的分析发现，用户的提及行为受其提及活动的语义和空间上下文因素的联合影响.据此，提出一个联合概率生成模型JUMBM（joint user mention behavior model），模拟用户空间关联提及活动的生成过程.通过对用户语义和空间上下文感知的提及行为进行统一建模，JUMBM能够同时发掘用户的移动模式、地理区域依赖的语义兴趣及其对应目标用户的地理聚集模式.此外，提出一种混合剪枝算法，加快推荐系统对在线top-k查询的响应速度.在大型真实数据集上的实验结果表明，所提方法在推荐有效性和推荐效率方面均优于对比方法.



Abstract:As a newly emerging social media user interactive service, mention mechanism is playing an important role in both information sharing and online social interacting. Researches on mention mechanism can provide us valuable resources to reveal the correlation between users' latent preferences and their explicit interacting behaviors and can be constructed as the data foundation for many applications such as information dissemination monitoring, business intelligence, and personalized recommendation. However, most of the previous works focused on the information diffusion aspect, lacking the in-depth study on its interaction attribute from the common users' perspective. This study aims to construct a recommendation system to automatically recommend target users for given social media posts based on the analysis and modeling of common users' mention behaviors. This study first analyzes two large-scale real-world datasets to explore the mention mechanism from the aspect of users' interactions and finds that, users' mention behaviors are impacted by both the semantic and the spatial context of their mention activities. Secondly, based on a unified definition of the joint semantic and spatial context-aware mention behavior, a joint latent probabilistic generative model named JUMBM (joint user mention behavior model) is built to simulate the generating process of users' mention activities. Specially, JUMBM is able to simultaneously capture users' movement patterns, geographical area-dependent semantic interests, and the geographical clustering patterns of the targets users. Besides, a hybrid pruning algorithm is proposed to achieve a fast high-dimensional retrieval and facilitate the online top-k query answering. Extensive experiments on real-world datasets demonstrate the significant superiority of the proposed approach over the baseline methods to make more effective and efficient recommendations.



PDF全文下载地址:

http://jos.org.cn/jos/article/pdf/5616