Study On Risk Analysis Of Societal Security Based On Machine Learning

Recently,severe societal security situation has been witnessed in China.It is necessary to conduct objective,effective and timely risk analysis against societal security incident.The qualitative risk analysis methods are costly and difficult to verify.Traditional semi-quantitative and quantitative risk analysis methods are often limited to fixed patterns and the flexibility is inadequate.But societal security risks are dynamic,highly uncertain,and sudden.So it is impossible to assess the societal security risks merely with traditional methods.In this study,methods based on machine learning are proposed,taking theft and terrorist attacks as examples of both conventional and unconventional societal security events,respectively.The main contents are as follows.(1)Analysis of risk sources and impact factors of conventional societal security events.Based on a variety of machine learning algorithms,prediction models of classification of larceny exconvict(first-offenders/recidivisms/recidivists)are established,extracting the duration,location,committing means and losses of the victim as the features.The results show that Random Forest has the best performance on prediction of larceny ex-convict,which has the highest F1-micro as 0.70 on training set(10-fold cross validation),and the F1 score for prediction of first offenders,recidivisms and recidivists are 0.73,0.75 and 0.76.The prediction methods of classification of larceny ex-convict are proposed,extracting static attributes of those such as gender,age,and education level etc.,and the appearances frequencies such as the frequencies for visiting Internet cafes during the month when those commit crimes,the frequencies for staying at hotels,and the frequencies for being checked by police as the features.The results show that Random Forest also has the best performance,which has with the highest F1-micro as 0.88.The frequencies for being checked by police,the age,the level of education and the distance from where they staying to where they commit crimes are the most significant feature to explain the types of larceny exconvict.(2)Analysis of the probability and consequences of unconventional societal security events.The terrorist attacks with low probability of occurrence and extremely serious consequences are selected as the research objects.The prediction methods for the numbers of terrorist attacks are studied,using the global terrorism database from 1970 to 2017.The results show that Random Forest and K-nearest neighbors perform best,whose goodness of fit R-square reaches 0.75 and 0.74,respectively in the test set.Furthermore,the generalization ability of Random Forest is stronger than K-nearest neighbors.Based on Prophet algorithm,the prediction method of the conditional probability of armed attack is studied.The results show that the average absolute error predicted by the model in 2016 and 2017 is 3.7%.Predicted seasonal trends indicate that the early February,mid-June and mid-October are the temporal hot spots of the armed attacks.What's more,the 12 features about time,place,attack intention,attack type and target type of the terrorist attack were conducted to predict the severity of the consequences of the terrorist attack.The results show that XGBoost has the best performance,and the F1-micro is 0.716 on training set and 0.72 on test set,respectively.