Mining Multidimensional Relationships Of Things In Cyber-Physical Space

With the rapid development and popularization of low-cost embedded sensing devices,wireless communication technologies and mobile computing technologies,the various tools and equipment used in daily life(e.g.,mobile phones,furniture,wallets and key chains)will possess sensing and communication ability.These physical things are heterogeneous in terms of function,type,manufacturer and other attributes,which will become the basic elements in cyber-physical space.The interconnection of massive heterogeneous things in the cyber-physical space gradually transforms people's work and living environment into an intelligent space,providing the ability to integrate the physical world and the cyber world,and also promote the emergence of many new applications(e.g.,things browsing and search,things semantic annotation and recommendation).One prerequisite step to implement these cyber-physical applications is to model and capture multi-dimensional relationships among heterogeneous things,including thing's relationship based on their attributes,thing's relationship based on their usage event as well as the relationship between things and users based on interaction behavior.Meanwhile,mining multi-dimensional relationships among massive heterogeneous things is also the key to build intelligent cyber-physical space.However,the interconnection of massive heterogeneous things also leads to implicit,dynamic and diverse characteristics of thing's relationship,which bring serious challenges to mine multi-dimensional relationships of heterogeneous things in cyber-physical space.This paper studies a set of technical frameworks for mining and analyzing multidimensional relationships of heterogeneous things in cyber-physical space,which not only can help people to acquire and organize the information of various things using a unified way,but also can be beneficial to a few cyber-physical applications(e.g.,things browsing and search,things semantic annotation and context-aware recommendation).To attack the challenges caused by the implicit,dynamic and diversity characteristics of thing's multi-dimensional relationships,the proposed frameworks mine thing's multi-dimensional relationships from three perspectives: 1)thing's relationship based on their attributes;2)thing's relationship based on their usage event and 3)the relationship between things and users based on interaction behavior.To indicate the effectiveness of mining thing's multi-dimensional relationships,this paper provides a real application scenario for each type of thing's relationship.Specifically,the contribution of this paper mainly includes the following aspects:? First of all,for mining thing's relationship caused by their usage event,this paper regards this kind of relationship as the hidden factor of thing's usage event with a latent variable model.In addition,this paper proposes a method for semantic labeling of things based on this kind of relationship mined from thing's usage event,which can automatically predict semantic labels for a given thing.The experimental results show that compared with the traditional semantic annotation models,modeling thing's relationship from their usage event can boost thing's semantic labeling and the improvement of labeling accuracy is more than20%.? Secondly,for mining thing's relationship caused by their attributes,this paper proposes a hierarchical relation model based on ontology.In the hierarchical relation model,the upper layer is some basic relationships(such as time,space,service and event),and the lower layer is the further subdivision and expansion of the relation of the upper layer.Based on this hierarchical relation model,this paper proposes a search model for things by considering user's context.The experiment results show the proposed search model can obtain higher user satisfaction.The reason is the proposed search model considers both user's search context and the relationship of things caused by their attributes.? Thirdly,for mining the relationship between people and things caused by their interaction behaviour in physical space,this paper proposes a hidden variable model to uncover this kind of relationship.Meanwhile,this paper proposed an indoor subarea localization based on graph matching to collect the interaction behaviour between users and things,which extracts the interaction behaviour from user-generated Wi Fi logs with passive crowdsourcing.This paper presents a store recommendation model based on user's preference from user-store relationship.The experimental results show that the proposed store recommendation model not only can improve the recommendation accuracy,but also provides richer recommended things.? Finally,this paper presents a thing's recommendation algorithm based on multimodal data set,which jointly considering the interaction between users and objects in physical space and cyber space.We implement a prototype of the proposed recommendation algorithm in two large shopping malls-physical store recommendation.Specifically,the store prototype recommendation system consists of three stages: 1)mining user-store preference relationship based on their interaction behavior in the physical space;2)mining store-aspect relationship based on the interaction behavior in the physical space between users and stores;3)utilizing a tripartite graph to capture both user-store preference relationship and store-aspect relationship,and performing store recommendation using a random walk propagation algorithm.The experimental results show that compared to the existing work,the performance of the proposed recommendation model can achieve more than 47% improvement.This paper focuses on three fundamental relationships among cyber-physical things and analyzes the extension and connotation of various relational mining models.We propose a set of methods and technical framework for analyzing various relationships among cyber-physical things,which can facilitate a few intelligent applications in cyber-physical space.