Research On Risk Assessment Of Prisoners Based On Machine Learning

Prisoners are an important and special group in society.Although the total population accounts for a small proportion,it has a great impact on the healthy and rapid development of the country,the harmony and stability of the society,and the safety of citizens’ personal and property.Therefore,the management and research of prisoners is a key and task.The daily behavior monitoring and risk assessment will help improve the management efficiency of the supervision site and ensure the safety of the supervision environment.In prison management,conventional risk assessments are mainly aimed at predicting and categorizing the risk levels and types of prisoners and the risk of recidivism.This thesis focuses on the risk level assessment.At present,in the research field of risk assessment,mature assessment techniques mainly include multiple forms of comprehensive-scale tools,correlation analysis and other mathematical-statistical methods,and binary logistic regression.Most of the prison risk assessment work mainly relies on the observation and evaluation of the police officers in each prison,and the judgment is based on the thesis quality table.There are problems such as strong subjectivity,poor real-time performance,and low accuracy.With the improvement of the inmate management system,it is also a big challenge to effectively process multi-dimensional and multi-type massive data.Aiming at this problem,this thesis mainly conducts research on the risk assessment of prisoners based on machine learning.The risk assessment is transformed into a multi-classification task.The multi-dimensional information fields and risk levels of prisoners are regarded as the feature set and label,respectively.Using machine learning and other related methods to build a risk assessment model for prisoners.First of all,sample information of prisoners are obtained in a pilot prison in Ningxia through online collection and offline collection.After preprocessing the original data,a structured data set of prisoners is constructed.Secondly,an integrated feature selection algorithm is proposed to mine the key feature fields that affect the danger of prisoners,delete redundant features and reduce the negative impact.Then,based on the machine learning classification algorithm,a risk assessment model for prisoners is constructed.Using Python to perform prediction and verification in a real sample set,the model can be directly applied to prison equipment for risk assessment and prediction of prisoners through the data interface.Finally,this thesis proposes a very high-risk early warning program,which sends out warning signals about a small number of very high-risk groups to provide help for prison management and control.The integrated feature selection method proposed in this thesis is suitable for the feature selection of structured data sets in classification tasks.This method combines data analysis of different labels based on a single feature and feature-label analysis.The importance of each feature is measured by calculating four indicators of different dimensions,and finally a feature subset for classification is formed,and the better part of the feature is selected for classification by sorting.So as to improve the accuracy of data classification.The main contributions of this thesis are as follows:(1)In the work of risk assessment and prediction of prisoners,a machine learning algorithm-based assessment model is proposed,which solves the problems of high subjectivity and poor real-time performance.(2)The association relationship between the data of different labels under a single feature is discovered.Under a single feature,the distribution of different labels of data is analyzed based on KDE.The probability density curve is used to calculate the misclassification probability and correlation between different labels of data,so as to measure the performance of the feature in the classification.(3)An integrated feature selection scheme based on multi-dimensional index fusion is proposed.It integrates the analysis of the relationship between features and labels and the relationship between the data of different labels under a single feature.Compared with the existing methods and results,the algorithm proposed in this thesis can effectively improve the classification accuracy of the same model.(4)Experiments were conducted on real data sets of pilot prisons.The results show that the risk assessment model of inmates based on integrated feature selection and machine learning methods proposed in this thesis is effective.The key factors affecting the danger of prisoners have been unearthed.The classification accuracy rate of the risk level reached 86.06%,which is 3.08%higher than the accuracy rate without feature selection.Under the same other experimental settings,the accuracy is increased by 3.82%compared with the information entropy and by 1.88%compared with the correlation feature selection respectively.Therefore,the current risk assessment work in prisons can be optimized to ensure the objectivity and accuracy of risk evaluation.