Research On Urban Crime Prediction Model Based On Space-time And Category-graph Fusion Network

Crime is one of the most complex social problems around the world,posing a major threat to human life and property.Predicting crime incidents in advance can be a great help in fighting against crime,e.g.,planning patrol routes for the police so as to efficiently utilize limited police resources and contain crime rates.During the past decade,crime prediction has drawn continuous attention from both academic and industrial communities.Although a plethora of methods have been proposed,most of the algorithms either perform prediction by leveraging linear or other oversimplified models with manually selected features or fail to fully explore the dynamic patterns in the crime data.This is because criminal activities are affected by many intricate factors from different aspects:(ⅰ)intra-region temporal correlations;(ⅱ)inter-region spatial correlations;(ⅲ)complex categorical dependency;(ⅳ)potential connections between external auxiliary data and criminal activities.In this thesis,we construct two novel urban crime prediction models based on deep learning to deal with the above challenges,i.e.,Crime Prediction Model based on Fine-Grained Spatial-Temporal Fusion Network(CrimeST)and Crime Prediction Model based on Space-Time and Category-Graph Fusion Network(CrimeSTC).Specifically,the Crime Prediction Model based on Fine-Grained Spatial-Temporal Fusion Network(CrimeST)captures the temporal and spatial correlations in criminal activities from multiple perspectives.First,we analyze the temporal and spatial factors of the criminal activities.Then,according to the analysis results and corresponding challenges,four components of the model are proposed: spatial dynamic fusion encoder(using local CNN to capture spatial correlations more granularly and applying dynamic feature attention mechanism to fuse information),temporal dynamic fusion encoder(using GRU and attention mechanism to extract fixed-period and short-term effects in a fine-grained manner and integrate latent information effectively),categorical encoder(using one-hot encoding to embed categorical information into the model)and joint training mechanism(combining dynamic representation and categorical information).Finally,extensive experiments are conducted on real-world data sets which clearly demonstrate the effectiveness of our model.In addition to spatial-temporal factors,criminal activities are closely related to many other factors which are conducive to enhancing prediction performance.Therefore,we improve CrimeST and propose the Crime Prediction Model based on Space-Time and Category-Graph Fusion Network(CrimeSTC)which can jointly learn the intricate spatial-temporal-categorical correlations hidden inside the crime and urban big data.The major improvements lie the following three aspects: spatial similarity calculation,external data embedding,and categorical dependency processing.First of all,we analyze the influencing factors of the criminal activities.Then,we optimize the fine-grained spatial-temporal fusion prediction model according to the analysis results where four components of the model are proposed,i.e.,self-improving spatial-temporal dynamic fusion encoder(improving the extraction of spatial similarity and using fully connected layers to process time-varying dynamic external data),static embedded encoder(handling the data that remain the same over time via fully connected layers),category-dependent encoder(modeling categorical information with a topological map and using GCN to extract categorical features),and joint training mechanism(concatenating dynamic and static representations as well as categorical information to forecast criminal activities).Finally,we perform experiments on real-world data sets to verify the effectiveness of the improvements.