| Name: | Description: | Size: | Format: | |
|---|---|---|---|---|
| 970.93 KB | Adobe PDF |
Advisor(s)
Abstract(s)
Sistemas de gestão de ofensores (Offender Management Systems) têm vindo a ser adotados cada vez mais em inúmeros países como Estados Unidos da América[1], Singapura[2]
Finlândia [3] entre outros países. A sua implementação permite que sistemas prisionais forneçam uma melhor experiência de reabilitação para os ofensores, uma monitorização mais
detalhada e o calculo de fatores de risco durante e após a sua sentença. Porém a literatura
existente descreve que no que toca ao calculo de risco de reincidência, estes sistemas tendem
a descriminar alguns grupos étnicos e favorecer outros. Esta dissertação tem como objetivo
em primeira instância, reunir e analisar resultados obtidos na literatura e posteriormente
comparar diferentes métodos de machine learning para este calculo de reincidência para
comparar resultados e combater a descriminação presente atualmente.
A analise efetuada nesta dissertação, comparou 3 modelos de machine learning desenvolvidos com 3 métodos diferentes, Adaboost Classifier, Logistic Regression e Random Forest
Classifier, cada um destes modelo visa o calculo de reincidência de um ofensor. Para a avaliação destes 3 modelos foram usadas as seguintes métricas Score, Precission, Recall e F1
Score e a comparação de características de entrada com maior importância na construção
do modelo. Foi também feita a analise de dois datasets, COMPAS Score dataset e NIJ Recidivism Challange Dataset, em que foram comparadas e analisadas as características dos
ofensores como a idade, género e etnia.
Offender Management Systems have been increasingly adopted in numerous countries such as the United States of America, Singapore, Finland and other countries. Their implementation allows prison systems to provide a better rehabilitative experience for offenders, more detailed monitoring and calculation of risk factors during and after their sentence. However, the existing literature describes that when it comes to calculating risk of recidivism, these systems tend to discriminate against some ethnic groups and favor others. This dissertation aims in the first instance, to gather and analyze results obtained in the literature and then compare different methods of machine learning for this recidivism calculation in order to compare results and combat the discrimination currently present. The analysis performed in this dissertation, compared 3 models of machine learning developed with 3 different methods, Adaboost Classifier, Logistic Regression and Random Forest Classifier, each of these models aims at the calculation of recidivism of an offender. For the evaluation of these 3 models the following metrics Score, Precission, Recall and F1 Score were used and the comparison of input features with greater importance in the construction of the model. Two datasets were also analized, COMPAS Score dataset and NIJ Recidivism Challange Dataset, where the characteristics of offenders such as age, gender and ethnicity were compared and analyzed.
Offender Management Systems have been increasingly adopted in numerous countries such as the United States of America, Singapore, Finland and other countries. Their implementation allows prison systems to provide a better rehabilitative experience for offenders, more detailed monitoring and calculation of risk factors during and after their sentence. However, the existing literature describes that when it comes to calculating risk of recidivism, these systems tend to discriminate against some ethnic groups and favor others. This dissertation aims in the first instance, to gather and analyze results obtained in the literature and then compare different methods of machine learning for this recidivism calculation in order to compare results and combat the discrimination currently present. The analysis performed in this dissertation, compared 3 models of machine learning developed with 3 different methods, Adaboost Classifier, Logistic Regression and Random Forest Classifier, each of these models aims at the calculation of recidivism of an offender. For the evaluation of these 3 models the following metrics Score, Precission, Recall and F1 Score were used and the comparison of input features with greater importance in the construction of the model. Two datasets were also analized, COMPAS Score dataset and NIJ Recidivism Challange Dataset, where the characteristics of offenders such as age, gender and ethnicity were compared and analyzed.
Description
Keywords
Adaboost
Classifier Algoritmos de Apoio à Decisão Compas Dataset Equidade Etnia Genero Gestão de Ofensores Idade Imparci-
Alidade Inteligência Artificial Justiça Logistic Regression Lsi-R Nij Recidivism
Challange Dataset Ofensores Random Forest Reincidência Risco Savry