Research On Anomaly Accident Prediction Via Multi-Source Heterogeneous Spatio-Temporal Data

In recent years,with the improvement of people’s attention to social security,the prediction of anomaly events that seriously threaten social security has also become one of the significant research issues.Anomaly events such as traffic crashes and crime events seriously threaten the safety of citizens’ lives and property.Accurate prediction of anomaly events is of great significance to maintain social security.For example,as for the government,the police force can be reasonably allocated to prevent crime and regulate traffic,and as for individuals,travel routes can be reasonably identified to reduce the risk of accidents.In the process of anomaly event prediction,the following three challenges need to be overcome:(1)Multi-dimensional heterogeneity of spatio-temporal data.The occurrence of anomaly events is often affected by numerous complex factors,such as time,weather,traffic flow,road conditions and POI,etc;(2)Dynamic variability of spatio-temporal data.The correlation and the dependence between anomaly events change dynamically as time goes by;(3)Sparsity of anomaly events.The data distribution is unbalanced,ie,the number of normal events is far greater than that of anomaly events.Previous work mainly focuses on utilizing fitting approaches with spatio-temporal dis-tribution or feature-based regression algorithms to solve this problem,which usually require strong parametric hypotheses and artificially designed features with poor scalability and pre-diction performances.Hence they are ineligible to apply to the actual scenarios.In order to tackle the complexity,dynamic variability of multi-source heterogeneous spatio-temporal data and sparsity of anomaly events,driven by the practical application of prediction results of anomaly events,this thesis proposes several spatio-temporal neural network models respec-tively to solve the above three problems and improve the prediction performance of anomaly events.The main innovative work and research results of this thesis are as follows:1.In order to tackle the sparsity of anomaly events and multi-source heterogeneity of spatio-temporal data,this thesis proposes an anomaly event prediction model based on joint static-dynamic spatio-temporal evolutionary learning.Through mining dynamic evolutionary fea-ture representations and static spatio-temporal correlations,multi-graphs are constructed by employing the inherent attribute similarities of nodes,and the graph convolution neural net-work based on attention mechanism is utilized to capture the spatial correlation and dependence among nodes to fuse heterogeneous features;the model proposes a conditional dynamic evolu-tion strategy,which further encodes the representation features of events over time by Update and Decay operations;by tuning the loss weight of different types of events and increasing the distance between related nodes,a loss function is designed to solve the problem of data imbalance,and the over-smoothing problem in the process of dynamic feature evolution is ef-fectively avoided;abundant experimental results on two large-scale real urban traffic accident data sets show that this model achieves better prediction performance and provides suggestions and references for urban road network regulation and traffic management.2.In order to capture the dynamic variability of anomaly events,the spatial correlation and the dependence of spatio-temporal data,a co-evolutionary graph convolution network spatio-temporal model for anomaly event inference based on attention mechanism is proposed in this thesis,which can reveal the evolution pattern and the interaction mode between dynamic events;this model embeds multi-source heterogeneous spatio-temporal data such as event be-havior,location and interaction information into event features,illustrates the dynamic process between events through the embedded features,and captures the interaction between behavior and location;by calculating the correlations between different events and different places dur-ing different time intervals,the edge weight is obtained and the attention matrix is generated;through random walk sampling,a graph convolution network with attention mechanism is used to aggregate behavior and location neighbor features over time,and the aggregated features are projected into multi-layer perceptron to predict the corresponding tasks.Abundance of experi-ments on the New York City crime event data set show that this model can predict the behavior type,location type and occurrence time of dynamic events with higher accuracy,lower time complexity and stronger scalability.According to the spatio-temporal data of relevant urban crimes,it can be used to predict the types of and the locations of crimes,and discover anomaly people effectively.