A carregar...
Publicação
Exploração de indicadores para identificar víeis em diferentes conjuntos de dados
| Nome: | Descrição: | Tamanho: | Formato: | |
|---|---|---|---|---|
| 2.45 MB | Adobe PDF |
Autores
Orientador(es)
Resumo(s)
Os conjuntos de dados frequentemente contêm vieses que prejudicam determinados grupos. Geralmente, é muito difícil eliminar totalmente esses vieses. No entanto, deve-se sempre tentar minimizá-los. O presente estudo tem como principal objetivo identificar a presença de viés em diferentes conjuntos de dados. Para validar a abordagem, foram realizados testes práticos em Python, com diferentes conjuntos de teste. Foram utilizados dois conjuntos de dados de contextos distintos. Aplicaram-se os testes estatísticos Qui-Quadrado ( 2) e ANOVA para identificar possíveis vieses, interpretando os resultados em função do nível de significância de 1% e dos testes de hipóteses. Além disso, aplicou-se a técnica de balanceamento de dados Random Oversampling, uma vez que os resultados continham vieses. A combinação dos testes estatísticos e da técnica de balanceamento permitiu analisar os resultados dos testes estatísticos antes e depois do balanceamento, demonstrando a eficácia das abordagens na deteção e mitigação de vieses, contribuindo para sistemas mais justos.
Datasets often contain biases that disadvantage certain groups. It is generally very difficult to eliminate these biases. However, one should always try to minimize them. The main objective of this study is to identify the presence of bias in different datasets. To validate the approach, practical tests were performed in Python with different test sets. Two datasets from distinct contexts were used. Chi-square ( 2) and ANOVA statistical tests were applied to identify possible biases, interpreting the results based on a significance level of 1% and hypothesis testing. In addition, the Random Oversampling data balancing technique was applied, as the results contained biases. The combination of statistical tests and the balancing technique allowed for the analysis of the statistical test results before and after balancing, demonstrating the effectiveness of the approaches in detecting and mitigating biases, contributing to fairer systems.
Datasets often contain biases that disadvantage certain groups. It is generally very difficult to eliminate these biases. However, one should always try to minimize them. The main objective of this study is to identify the presence of bias in different datasets. To validate the approach, practical tests were performed in Python with different test sets. Two datasets from distinct contexts were used. Chi-square ( 2) and ANOVA statistical tests were applied to identify possible biases, interpreting the results based on a significance level of 1% and hypothesis testing. In addition, the Random Oversampling data balancing technique was applied, as the results contained biases. The combination of statistical tests and the balancing technique allowed for the analysis of the statistical test results before and after balancing, demonstrating the effectiveness of the approaches in detecting and mitigating biases, contributing to fairer systems.
Descrição
Palavras-chave
ANOVA Balanceamento de dados Inteligência artificial Qui-Quadrado Viés algorítmico