Research On Urban Sparse Anomaly Prediction Method Based On Joint Spatio-Temporal Attention Graph Network

Urban anomalies may pose tremendous threats to public safety and stability if not handled properly.Urban anomaly predictions,such as traffic accident risk prediction and crime prediction,are of vital importance to smart city security and maintenance.There are already many researchers working on the problem of urban anomaly prediction based on big data and artificial intelligence.However,urban anomalies are a kind of temporal facts with complex causes and many dimensions of influence factors.Besides,compared with other temporal facts,urban anomalies are rare events with zero-inflated issue.To address the above challenges,this thesis proposes two solutions based on joint spatio-temporal graph networks for the urban anomaly prediction problem with sparse data.Firstly,this thesis proposes a zero-inflated urban anomaly prediction method based on the spatio-temporal attention graph.Specifically,in the spatio-temporal dependency module,the spatio-temporal dependency layer employs a dynamic spatio-temporal attention module to capture cross-spatial and long-term temporal correlations from onehop neighbors.Then,the module iterates the layer to spread that influence to multi-hop neighbors.In order to alleviate the zero-inflated issue,a multi-task prediction framework is designed to simulate the exposure process and count process of the occurrence of urban anomalies.Meanwhile,a customized loss function is proposed to further alleviate the zero-inflated issue.Finally,this thesis perform complete experiments,including comparison experiments,ablation experiments and hyperparameter learning,on four real-world datasets in two application scenarios(i.e.,crime prediction and traffic accident risk prediction)to demonstrate the improved prediction capability and robustness of the method.Secondly,this thesis proposes a urban anomaly prediction method based on the multidimensional implicit spatio-temporal graph.This method extends the zero-inflated urban anomaly prediction method based on the spatio-temporal attention graph by extending the spatio-temporal dependency module into the spatio-temporal and semantic dependency module.Specifically,in the spatio-temporal and semantic dependency module,this method designs a multi-head semantic self-attention module to capture dependencies from different categories;a time-aware recurrent module to capture correlations in the neighbor temporal dimension;and a global spatial hypergraph learning module to model implicit correlations in the semantic spatial dimension.Through extensive experiments on four real-world datasets in two representative application scenarios(i.e.,crime prediction and traffic accident risk prediction),including comparison experiments,ablation experiments and hyperparameter learning,this thesis demonstrates the improved prediction capability,robustness and generality of the method.Finally,more detailed experiments and analysis of the sparsity problem are conducted in this thesis.The results show that the proposed methods in this thesis can well alleviate the data sparsity issue in urban anomaly prediction research.