Research And Implementation Of Few-shot Urban Abnormal Prediction

With the acceleration of urbanization,crimes,traffic accidents and other abnormal events occur in cities from time to time.Public safety has become the focus of attention of city managers all over the world.Urban abnormal events usually refer to unusual events occurring in the urban environment.If not handled in time,they may pose a great threat to the safety of people’s personal property and the development of society.If abnormal events can be predicted in advance,city managers will have more time to formulate corresponding measures for abnormal situations,thus effectively reducing potential risks.Therefore,the research of urban abnormal event prediction technology is of great value to ensure public security.The urban abnormal prediction refers to the identification,analysis and prediction of future abnormal events by using the historical records of abnormal events,so as to realize the advance feedback of information and reduce the loss caused by potential abnormal events to the greatest extent.However,urban anomalies themselves rarely occur,and due to the high cost of privacy protection,data collection management and maintenance,their historical record is very limited,and their prediction generally faces the problem of sparsity.Existing research methods still have some shortcomings: On the one hand,existing research tends to use multi-source city data to enrich context information for prediction of common types of abnormal events such as theft.However,for robbery,arson and other few-shot data types,this method is difficult to increase the number of samples,and cannot adapt to the few-shot data prediction scenarios with extremely sparse data.On the other hand,due to the influence of data sparsity,the distribution of urban abnormal events is discrete and unbalanced in the spatio-temporal dimension,which does not conform to the basic assumption of continuous distribution in the temporal sequence of traditional temporal forecasting models.Therefore,in view of the data characteristics of few-short city anomaly data,this study designed a prediction model that can effectively alleviate the problems of sparsity and discrete distribution imbalance,and did experiments in real data sets.The specific work is as follows:Firstly,to solve the problem of direct modeling of few-shot data is prone to overfitting,this study proposes a cross-city abnormal data augmentation method based on transfer learning to enrich the number of training samples.In order to achieve better migration effect,this thesis adaptively matches different regional granularity for each type of abnormal event,considering the development difference between cities and the distribution difference of different types in different cities.Using the enhanced anomaly information for training,the prediction performance improved by 12%.Secondly,to solve the problem of discrete and unbalanced distribution of few-shot data in spatio-temporal dimension,a prediction model of urban abnormal events based on Neural ODE is proposed to process discrete and limited historical abnormal records for continuous multi-slot abnormal events prediction.Using real data sets,the prediction performance of the proposed model is improved by 25.9%.Thirdly,to meet the needs of users to predict different types of urban abnormal events with different granularity(such as path planning,block patrol,etc.),an interactive urban abnormal event prediction prototype system was designed and implemented.The offline part of the system is responsible for data management and model training,and the online part of the system is responsible for visualizing the prediction results of urban abnormal areas that meet the personalized needs of users,and the accuracy of the prediction performance is verified in the system performance.To sum up,this thesis aims at the prediction of abnormal events in cities under social public security scenarios,and proposes corresponding data augmentation methods and anomaly prediction models according to the characteristics of abnormal event data in cities with few-shot types.Based on the theoretical research results,an interactive prototype system for urban abnormal event prediction is designed and implemented.The theoretical mechanism and prototype system proposed in this thesis contribute to the construction of smart city security services,which can be further applied in many fields such as city perception and safe travel.