Scikit-learn: um estudo sobre aprendizado de máquina em python

Abstract

Este trabalho buscou demonstrar como a técnica de Machine Learning pode ser aplicada por meio de algoritmos classificadores disponível em Python por meio da Biblioteca Scikit-learn. Para isso, os conjuntos de dados Iris e Titanic foram utilizados. Os algoritmos utilizados foram: Random Forest, Decison Tree, Extra Trees, Adaboost e K-Nearest Neighbors. Para comparar os resultados obtidos, foi utilizada a métrica acurácia, que representa a porcentagem dos valores preditos de forma correta em relação ao valor real, e essa métrica foi apurada com dois métodos: accuracy_score e cross_val_score. Com isso, foi constatado que no datasetTitanic com o método accuracy_score, o melhor modelo foi o AdaBoost, no entanto,com o método cross_val_score houve um empate entre o Extra Trees e o Random Forest. Já no datasetIris, com o método accuracy_score, os melhores desempenhos foram dos modelos Adaboost e K-Nearest Neighbors e, com o método cross_val_score, o melhor desempenho foi para o algoritmo K-Nearest Neighbors.

This work focused on demonstrating how the Machine Learningtechnique can be applied using classifying algorithms available in Python through the Scikit-learn Library. For this, the Iris and Titanic data sets were used. The used algorithms were: Random Forest, Decison Tree, Extra Trees, Adaboost and K-Nearest Neighbors. To compare the obtained results, the accuracy metric was used, which represents the percentage of the values correctly predicted in relation to the real value, and this metric was calculated using two methods; accuracy_score and cross_val_score. With that it was found that in the Titanic dataset with the accuracy_score method the best model was Adaboost, however,with the cross_val_score method there was a tie between Extra Trees and the random Forest. In the Íris dataset, with the accuracy_score method, the best performances were from the Adaboost and K-Nearest Neighbors models and, with the cross_val_score method, the best performance was for the K-Nearest Neighbors algorithm.