【作者】 陈明；

【导师】 李文中； 陆桑璐；

【作者基本信息】 南京大学 ， 计算机软件与理论， 2019， 博士

【摘要】 当前,基于位置的服务正在获得越来越广泛的应用,然而传统的仅依赖定位设备的位置服务所能提供的信息是相对有限的。随着移动互联网、智能手机技术的发展,移动社交网络中不仅包括了丰富的用户数据,如好友、兴趣、社交数据等,而且沉淀了大量的用户位置轨迹数据。在移动社交网络中的进行位置轨迹挖掘,有着重要的意义,它可以应用于基于位置服务、挖掘人类活动规律与行为特征,还可以应用于智能交通、智慧旅游、环境监测、能源消耗等城市计算项目。如何充分挖掘用户的位置轨迹,并结合社交网络的多元信息为用户提供更加便捷、准确、安全的位置服务,是当前移动社交网络研究的重要问题之一面向移动社交网络中位置轨迹的挖掘及应用,我们希望建立基于位置服务中高效且精确的用户兴趣信息推荐、聚类机制,最大限度地将用户的兴趣上下文场景反映到模型之中,从而提高基于用户特性的推荐、聚类算法的效果;同时,建立用户的位置隐私保护机制,以达到用户放心使用移动社交网络的目的。具体而言,我们在基于移动社交网络用户位置轨迹数据的挖掘方面,开展了一系列研究工作,主要包括基于位置服务中的位置推荐、社群聚类和隐私保护。在基于位置服务的位置推荐工作中,下一跳兴趣点推荐(Next POI Recommendation)是为用户推荐他愿意去的下一个感兴趣的位置,难点在于需要充分利用移动社交网络数据(兴趣、位置、好友关系等),结合推荐技术,以达到为用户推荐合适位置的目的。在基于位置服务的社群聚类工作中,建立基于地理位置的社群,对人们深入理解基于移动社交网络的拓扑结构和社群主题等,具有重要的意义和研究价值。难点在于移动社交网络拥有用户、地理位置等多个实体,以及用户关系、用户兴趣关系、用户位置关系等多个关系,在这种复杂网络中挖掘地理位置社群,融合多模异构的实体和关系是地理位置社群聚类分析的关键。此外,位置隐私保护(Location Privacy Preserving)是基于位置服务中一个日益令人关切的问题。在基于位置服务的位置隐私保护工作中,只有用户感到隐私会被妥善保护,其才更愿意使用移动社交网络。位置隐私保护工作的难点在于用户提供的信息和获得的位置信息服务是一对矛盾,然而,用户提供的数据越少,其能获得的位置服务质量越差。针对上述问题,我们从以下三个方面开展工作:第一,针对基于位置服务中的位置推荐问题,传统算法未能同时考虑用户兴趣、位置轨迹时间依赖关系和个性化位置信息,我们提出了一种基于长短期记忆循环神经网络(LSTM)的兴趣相关下一跳POI推荐方法。我们通过实证观察到用户兴趣、位置轨迹时间依赖关系、上下文场景对用户位置推荐效果有很大的影响。我们通过分析挖掘隐藏用户位置轨迹信息,有效地发掘用户的兴趣,结合上下文场景信息,可以准确地为用户预测推荐下一跳位置。这项工作中,基于真实数据集的实验证明了我们方法的有效性。第二,针对基于位置服务中的社群聚类问题,在移动社交网络中挖掘地理位置社群的难点在于融合多模异构的实体和关系,我们提出一个基于地理位置社群聚类的算法。该算法使用主题模型抽取用户兴趣,使用深度学习技术融合了用户关系、用户兴趣关系、用户位置关系,来确定社群划分。实验结果验证了我们算法的有效性,并展示了所划分的社群在用户特性上的内聚性。第三,针对基于位置服务中的位置隐私保护问题,传统算法基于拉取策略有着难以避免不同级别的LBS服务器或代理被破解的缺点,我们提出使用分布式缓存推送来保护位置隐私。其基本思想是应用分布式缓存代理来存储最流行的与位置相关的数据,并主动将数据推送给用户。如果用户所需的数据可从缓存中获得,则不需要发送基于位置的查询,从而保护了用户的隐私。我们提出的名为LPPS的基于推送的位置隐私保护方案,引入了分布式缓存层来存储热门的与前位置相关的数据,并将其推送到移动用户。我们提出了缓存代理部署和缓存推送的策略,以实现位置隐私的k匿名性。在缓存不命中的情况下,提出了缓存替换和更新策略,以挖掘大量从假反馈索引中隐藏的真正的热门场所。基于轨迹数据的仿真实验表明,该方案有较高的服务覆盖率、较好的缓存命中率以及较低的通信开销。更多还原

【Abstract】 Nowadays,location-based services(LBSs)become more and more popular.However,the information provided by conventional LBS that solely relies on positioning equipment is relatively limited.With the development of Mobile Internet and smart phone technology,mobile social networks not only contain user big data,such as friends,interests,social data,etc.,but also precipitate a large amount of user location trajectory data.Location trajectory mining in mobile social networks has important significance.It can be applied to location-based services,mining human activity patterns and behavioral characteristics,and can also be applied to urban computing projects such as intelligent transportation,smart tourism,environmental monitoring,and energy consumption.How to fully exploit the user’s location trajectory and the multiple information of the social network to provide users with more convenient,accurate and secure location services is one of the most important issues in the current mobile social network research.Aiming at location trajectory mining and its application in mobile social networks,we wish to establish an efficient and accurate user interest information recommendation and clustering mechanism in LBS to incorporate the user’s interest context into the model,xwhich can improve the recommendation and clustering performance based on user characteristics.At the same time,based on the user’s location privacy preservation scheme,we hope to achieve the purpose of guaranteeing user’s location privacy in mobile social network with confidence.Specifically,we focus on a number of topics in location-based services(LBS),including location recommendation,community detection and privacy protection.In location recommendation,the topic of“Next POI Recommendation "refers to recommend the next point of interest to the users.The difficulty is to make full use of mobile social network data(interest,location,friends relationship,etc.),combining them with recommendation techniques to achieve the purpose of recommending a proper location to the user.In location-based community detection,finding communities based on location interests has significant research value for people to deeply understand the topology and community topics in mobile social networks.The difficulty lies in the fact that mobile social networks have multiple entities such as users and locations,and multiple relationships such as user relationships、user interest relationships and user location relationships.In this complex network,integrating multi-mode heterogeneous entities and relationships is the key of location community detection.In addition,preserving location privacy is a growing concern in location-based services.Only if users’privacy be properly preserved,they are willing to use mobile social networks.The difficulty of location privacy protection is that the information provided by the user and the obtained location information service are contradictory.The less data provided by the user,the worse the quality of the location service that can be obtained.In response to the above problems,our work focuses on the following three aspects:Firstly,for location recommendation in LBS,traditional algorithms failed to capture user interest,location trajectory time dependence and personalized location information at the same time.We propose a recurrent neural network(LSTM)based next POI recommendation method.We empirically observe that user interest,location trajectory,time dependence,and context scenarios have a great impact on user’s location recommendation.We analyze and mine the hidden information behind user location trajectory,effectively explore the user’s interest,and combine the context information to effectively recommend the next POI for the user.In this work,experiments based on real datasets demonstrate the effectiveness of the proposed approach.Secondly,for community detection in LBS,the difficulty in mining location community in mobile social network lies in the fusion of multi-mode heterogeneous entities and relationships.We propose a deep learning algorithm for location interest community detection.The algorithm uses the topic model to extract user interests,and uses deep learning technology to integrate user relationships,user interest relationships,and user location relationships to detect community.Experimental results verify the effectiveness of the proposed algorithm and demonstrate the cohesiveness of the divided communities in user characteristics.Thirdly,for location privacy preservation in LBS,the traditional algorithms based on the pull strategy had the difficulty to avoid different levels of LBS servers or agents being compromised.We propose to use distributed cache pushing to protect location privacy.The basic idea is to apply distributed caching proxies to store the most popular location-related data and proactively push the data to the user.If the data required by the user is available from the cache,the user does not need to send a location-based query,thereby protecting the privacy of the user.The proposed push-based location privacy protection scheme called LPPS introduces a distributed caching layer to store popular data related to the current location and push it to mobile users.We propose a caching proxy deployment and cache push strategy to achieve k-anonymity of location privacy.In the case of cache misses,we propose a cache replacement and updating strategy to mine the real hot spots hidden in a large number of fake feedback indexes.Simulations based on trajectory data show that the proposed scheme has higher service coverage,better cache hit ratio and lower communication overhead compared to conventional approaches.更多还原