Avaliação de um modelo de previsão utilizando métricas estatísticas: um estudo de caso

Abstract

This work presents an analysis and evaluation of regression models through various statistical metrics. The analysis of regression models in data science and the need for a thorough evaluation of model performance, highlighting the relevance of performance metrics, is still an area where there are few studies in the state of the art. For the development of this analysis, the dataset of airline passengers was considered, as well as the evaluation of statistical metrics for the linear classifier. In this context, metrics such as Mean Squared Error (MSE), Root Mean Square Error (RMSE), Mean Absolute Error (MAE), Mean Absolute Relative Error (MARE), Relative Absolute Error (RAE), Mean Absolute Percentage Error (MAPE), Median Absolute Error (MeAE), Geometric Mean Absolute Error (GMAE), Sum of Squared Errors (SSE), Geometric Root Mean Squared Error (GRMSE), Kullback-Leibler divergence, Cauchy-Schwarz divergence, Bhattacharyya distance, Hellinger distance, Manhattan distance, Euclidean distance, Neyman Chi-Square distance, and Jensen-Shannon divergence (JS) were considered, applied to the linear regression classifier. Thus, the challenge of evaluating and analyzing statistical metrics in data analysis is highlighted, emphasizing that the traditional approach based on Euclidean distance is not always effective. This diversity of metrics, based on information theory, allows for the adjustment of mathematical models to capture specific nuances of the data, resulting in more precise analyses and informed decisions. Computational experiments revealed that the Hellinger distance metric outperformed other metrics that adopt a methodology with values ranging from 0 to infinity. The Cauchy-Schwarz metric also stood out on a scale of values ranging from 0 to 1.