Study Of Location Recommendation And Regional Demand Analysis For Urban Facility Based On Multi-Source Data Fusion

With the advancement of urban sensors,Internet of Things and social networks,the cross-disciplinary study of urban computing has developed rapidly.The recommendation of candidate locations for establishing service facilities is one of the most important fields in urban computing,which could support the convenience lives to residents,improve the service efficiencies and competitiveness of facilities,as well as alleviate the pressures on both urban traffic and living environment,etc.In addition,the characteristics of urban regions has varied over time in the process of urbanization,then identifying the POI demand in different regions can provide scientific decision-making for urban planning.However,current studies on location recommending and regional demanding of facilities mainly use the macro datasets of city itself,and lack of analysis from the perspective of micro datasets such as spatio-temporal trajectories and social networks made by residents.There(?),in this thesis.we make full use of multi-source perception datasets of urban residents,then integrate them effectively and efficiently to obtain complete and accurate data representations,which are adopted to support the study of location recommending and regional POI demand modeling.Specifically,the contents of our research include the following four parts.(1)We propose a multi-source heterogeneous data fusion model based on user similarity linkage.The existed fusion approaches do not comprehensively recognize the spatiotemporal characteristics and social relationship structures of users,which may lead to lower efficiency and inferior accuracy.Consequently,we devise a fusion model based on multiple datasets of users,e.g.,spatio-temporal trajectories,POI check-ins,and online social relationships.In order to improve fusion efficiency,both the time-period-based temporal index and urban-gird-based spatial index are adopted simultaneously.By calculating the spatiotemporal activity probabilities over indexes,the similarity matching spaces are then pruned.In order to improve fusion effectiveness,the relative entropy and kernel density estimation are introduced to assess the similarities of spatio-temporal features,as well as the pearson correlation coefficient is used to calculate the structure similarities in social relationships.Finally,according to the ultimate matching scores that are computed by the linearly integrated three previous kinds of similarity features,the multi-source datasets are fused with the help of the corresponding user linkages.Experimental results show that our approach is superior to baseline algorithms in terms of both efficiency and effectiveness.(2)We propose a candidate location recommending model for urban facility based on characteristic regression.The current work of location recommendation ignores the fact that urban residents are dynamically serviced entities,as well as does not comprehensively evaluate the influences in terms of their social relationships and activity time periods.Therefore,we dedicate to the analyses of those ignored factors,and simultaneously recognize both the service capacities and categories of facilities that are to be built.Specifically,by taking advantage of quantitative schemes of location popularity,we not only capture the spatiotemporal characteristics of residents from historical trajectories,but also calculate the social interactions from their social relationships and POI check-ins.In order to improve query efficiency,a B-tree-liked index is proposed to collect both customers and candidate locations into clusters on different tree levels.Based on the indexes,a heuristic greedy algorithm is put forward to obtain the final recommended results.Experiments on two real-world datasets show that the proposed model not only improves computational efficiencies and recommendation accuracy against baselines,but has better recommending interpretability.(3)We propose an incremental location recommending model for urban facility based on graph neural network.The regression-based approach is incompatible with the spatiotemporal incremental recommendation of candidate locations,since there are some of the congener facilities that have already been established in urban area.To this end,we propose a semi-supervised recommending model based on graph neural network and recurrent neural network.Firstly,by taking the locations as the solely centers,the location correlation graph is constructed,according to that the whole urban space is divided into neighborhood regions and the corresponding characteristics are extracted.Next,we introduce a contextual graph convolution module to represent spatial correlations among locations,where both the local and global spatial correlations are captured by two convolutional blocks based on multi-head attention mechanism,respectively.Subsequently,a recurrent neural network is adopted to illustrate the dynamic temporal dependencies inside of and between locations over time periods.And then,a entropy mechanism is represented to fuse the missing location popularity that is estimated from both temporal and spatial domains,which is afterwards sent to the recurrent neural network again to obtain the final popularity for recommending.Experimental results and theoretical analyses based on two real datasets show that the recommendation effects of the proposed model are superior to baselines.(4)We propose a predicting model for regional facility demand based on variational graph auto-encoder.The regional facility demand modeling is not only affected by the dynamically varieties of hotspot regions,POI facilities and residents in urban space,but faced with the realities of data sparsity and data uncertainty.To solve these problems,first we devise a two-level administrative partition scheme to divide urban space into spatially regions,then the representative features are extracted from each region,as well as the regional spatial attribute graph are formed.Meanwhile,a bayesian inference algorithm is proposed to predict the POI visit probabilities in the traveling destinations of residents.Next,a variational graph auto-encoder is adopted to map the original characteristics of graph nodes to the latent spaces,which is used to obtain the exactly representations of probability distributions.Then a spatial correlation module is put forward to capture the global correlations of regions and integrate their corresponding features sampled.Finally,a multi-layer perceptron module is introduced to predict the probability distributions of demands from two perspectives of facility category and POI facility.The experimental results show that the proposed model has better performances in effectiveness of demand predicting than baseline algorithms.In summary,based on the data fusion of multi-source heterogeneous information of urban residents,we have studied the problems of candidate location recommendation of POI facilities and regional POI demand modeling in urban planning,then proposed several recommending models,as well as proved the efficiencies and effectiveness of the proposed algorithms through theories and experiments.