Research On Urban Multi-source Heterogenous Data Fusion Methods And Applications

The rapid development of city brings people a modern life,but also brings many problems and new demands.Urban big data collected by the large number of sensors deployed in city offers the possibility to address these challenges.It is of great significance to effectively apply urban big data to solve urban problems and improve people’s life experience.This is also an important technical mean to build a smart city.In the research of urban computing,important urban big data includes temporal data,spatial data and knowledge graph.However,the trend analysis of temporal data,the feature association of spatial data,the multi-semantic embedding of knowledge graph,and the collaborative representation of these data bring great challenges to solve these problems.Therefore,the main goal of this dissertation is to study how to efficiently fuse these important urban big data in different application scenarios.Finally,this dissertation proposes the temporal data fusion method with symbolic aggregate approximation,the spatial data fusion method based on deep reinforcement learning,the cross-domain knowledge graph fusion method and the spatio-temporal data and knowledge graph fusion method.This achieves to explore novel urban multi-source heterogeneous data fusion techniques and practical applications in the city.(1)Urban temporal data fusion method and its application : Research on the symbolic aggregate approximation mapping method and decision fusion model of data.It is also applied to detect dangerous driving behaviors.First,the auxiliary data information is fused by the heterogeneous decision fusion method with the temporal data.Then,the initial fusion result is symbolized by the symbolic aggregate approximation method.Moreover,the correlation information between temporal data is extracted by the average fusion method and the homogeneous decision fusion method.Finally,a novel data fusion method based on symbolic aggregate approximation is designed to fuse temporal data.Experiments on traffic sensor dataset and weather dataset show that the proposed method can detect dangerous driving behaviors more effectively than baseline methods.(2)Urban spatial data fusion method and its application : Investigation of the deep reinforcement learning-based spatial data fusion network.It is also applied to optimal schedule recommend.First,a novel deep learning fusion network is designed to fuse multiple factors affecting schedule recommendation.Then,the basic elements of the reinforcement learning framework in the schedule recommendation scenario are defined.Finally,the parameters of the deep learning fusion network is learned by the policy gradient,and a deep network that can optimally fuse spatial data is obtained.Experimental results based on Chengdu travel dataset and Beijing trajectory dataset show that the proposed method is reasonable and effective.(3)Urban knowledge graph fusion method and its application : Research on cross-domain knowledge graph fusion method,and apply it to multi-domain item-item(I2I)recommendation.First,the knowledge graph of each domain is initialized by the classical knowledge graph embedding method.Then,a cross-domain knowledge graph crossembedding method is proposed to effectively interact with all entities and relations in multiple domains.Finally,a novel multi-domain I2 I recommendation method is designed based on the embedded knowledge graph.Experimental results based on public dataset FB15K-237 and cross-domain knowledge graph dataset show that the proposed knowledge graph embedding method and multi-domain I2 I recommendation algorithm achieve better link prediction and multi-domain recommendation results,respectively.(4)Urban spatio-temporal data and knowledge graph fusion method and its application :Investigation of urban spatio-temporal data and knowledge graph fusion model based on deep learning and knowledge graph embedding.It is also applied to mining urban traffic pattern.First,multiple auto-encoders are constructed to extract region features and traffic flow features.Then,an attention mechanism-based fusion method is designed to fuse regional features and traffic flow patterns.Finally,the construction of knowledge triples of traffic flow patterns is realized based on the translation distance method.The experimental results based on the Chengdu map,Chengdu Didi order dataset and the POI dataset show that the proposed model can effectively mine the knowledge of traffic flow patterns.In addition,visual analysis is performed by classifying regions(entities)and clustering traffic patterns(relations)between regions to demonstrate the application based on regional traffic patterns.