Aprendizado de máquina aplicado a um estudo de caso de classificação de qualidade de maçãs

Abstract

The use of machine learning (ML) has become prominent for its significant contributions across various fields, including Chemical Engineering, by enabling the analysis of large data volumes and extracting complex patterns. Classification models, in particular, have been widely applied to predict outcomes in industrial processes, allowing decision automation and efficiency improvements. This study focused on applying different ensemble techniques to decision tree models, aiming to investigate the performance of these approaches in classification problems. ensemble techniques, such as bagging and boosting, combine multiple models to improve overall performance, especially in scenarios where individual models may have limitations in terms of accuracy and robustness. The research aimed to evaluate the performance of ensemble techniques compared to standalone models, considering performance metrics such as accuracy, precision, recall, and F1-score. Additionally, model interpretability techniques, such as SHAP (Shapley Additive Explanations), were used to understand the impact of each variable on predictions and ensure that the results were interpretable and applicable to industrial contexts. Interpretability is a critical issue, especially in fields like Chemical Engineering, where understanding the model’s decisions can be as important as its accuracy. The results showed that both ensemble techniques, bagging and boosting, performed well, with high accuracy and robustness in most tasks. However, boosting exhibited a superior ability to adjust errors, effectively minimizing biases in the data. This technique proved particularly efficient in improving performance in cases where standalone decision tree models struggled with outliers or complex data patterns. The conclusions of this study reinforce the importance of exploring different machine learning techniques and evaluating their performance carefully, considering not only traditional metrics but also the interpretability of the models. Techniques like boosting offer promising solutions for complex problems, ensuring that models not only perform well but can also be understood and practically applied in real-world environments.