Research On Location Prediction And Recommendation Methods Of Mobile Objects Based On Spatio-temporal Context

With the rapid development of mobile Internet,positioning technology,wireless communication technology and smart mobile devices,various kinds of moving objects can be effectively located and tracked,which generates a huge amount of spatio-temporal trajectory data of moving objects and greatly contributes to the development of location-based social networks.The research of using these spatio-temporal trajectory data for location prediction and service recommendation of moving objects is an important direction for human mobility in the era of big data,which is also an significant research topic of common interest in the fields of Geographic Information Science,Computer Science,Sociology and Complexity Science.This research has important meaning in learning value and practice application in location social network services,intelligent traffic management,intelligent police,mobile computing,environmental monitoring,disaster management and smart city construction.Along with the development of big data,artificial intelligence,high-performance computing and other technologies,new opportunities have been brought into the research of human mobility modeling and prediction problems.However,mining large-scale trajectory data for the movement behavior of moving objects faces many challenges,including data sparseness,preference heterogeneity,spatiotemporal heterogeneity,long-term dependence and model efficiency.Addressing these challenges,this study searches for the novel way from two aspects including point of interest(POI)recommendation and vehicle location prediction.Several research works have been done with spatio-temporal trajectory data as the main research object and the main research goals including location prediction and recommended location services for moving objects.The main research results are as follows:(1)Due to limitations of traditional matrix factorization-based methods of POI recommendation using simple/fixed inner products to estimate complex user-POI interactions in low-dimensional latent space and the data sparsity problem of the user-POI matrix,This paper proposed a POI recommendation method based on joint geo-sequential preference and distance metric factorization model considering the combined role of users,POI,interest transfer patterns,and spatial context in a POI recommendation task.Firstly,the metric vector space is decomposed to learn the positions of users and POI in the metric space,and the Euclidean distance is used to replace the inner product representation of the matrix factorization method,then the generalized preferences of users are modeled by using the Euclidean distance of users and POI.Secondly,considering the sequence transfer characteristics of the user's interest,the Euclidean distances of consecutive check-in locations are used to model the user's interest transfer pattern.Then,the two types of distances are linearly fused,and the geographical distances are constrained as a weighting factor to construct a unified framework of POI recommendation that fuses general preferences,interest transfer patterns and geographic influences.Finally,a large number of experiments are performed on real datasets,which proves the effectiveness and superiority of the method.(2)The traditional shallow POI recommendation models are difficult to effectively integrate the spatiotemporal context information and model users' personalized differences in space and time.Existing researches fails to accurately model the long-term dependencies and capture the user's main behavioral intent.Therefore,this paper proposed a POI recommendation method based on spatiotemporal Gated Recurrent Unit and attention mechanisms.First,a spatio-temporal Gated Recurrent Unit network is constructed by combining continuous geographic distance information and time interval information of user movement and modeling the user's sequence preferences and personalized spatio-temporal preferences simultaneously.In addition,the structure of the recurrent neural network can be better learning long-term dependent.Then,learn the main behavioral intentions of user movement by introducing an attention model.The experimental results show that the method outperforms state-of-the-art POI recommendation methods in accuracy.(3)Although the recurrent neural network has been successfully applied in the POI recommendation,the calculation of the recurrent structure depends on the hidden state of the whole time sequence,which is difficult to improve the efficiency of the model effectively and learn the high-order feature interactions of the spatial dimension adequately.Therefore,this paper proposed a POI recommendation method based on spatiotemporal dilated convolutional generative network.First,a simple and efficient conditional generation model is introduced and the user mobile sequence is modeled by using a stackable dilated causal convolutional network and residual block structure that can be calculated in parallel.Then,the user's personalized spatial preference is modeled by using continuous geographic distance information and the two types of temporal periodic patterns are integrated to model the user's personalized time preference.The experimental results show that the method has obvious advantages in accuracy and efficiency compared with several types of state-of-the-art POI recommendation methods.(4)In order to solve the problem of long-term dependence of the modeling trajectory sequence in the vehicle location prediction task and the problem of low efficiency of the existing methods based on recurrent neural network and convolutional neural network,this paper proposed a vehicles moving location prediction method based on hierarchical attention mechanism and dilated convolutional generative network.First,we establish a refined geographic grid representation of large-scale vehicle trajectory data,obtain grid-coded sequences of moving trajectories,and then replace the original trajectory data as input to the prediction model.Next,a simple and efficient generative model is introduced to learn the deep-level important features of the vehicle trajectory and the spatiotemporal movement laws by integrating the hierarchical attention mechanism,dilated causal convolutional network and the residual block structure.The experimental results show that the method not only effectively solves the problem of long-term sequence dependences and obtains higher prediction accuracy,but also the operating efficiency of the model is greatly improved.The research results in this paper can not only provide novel ideas and technical reserves for the location prediction and service recommendation of moving objects,but also further enrich the existing basic theories of spatio-temporal data mining and machine learning,which have great theoretical significance and practical value.There are 46 figures,18 tables,and 226 references in this dissertation.