Comparativa entre regresión logística ordinal, redes neuronales artificiales y Gradient boosting; en la predicción de la satisfacción laboral en Ecuador

Abstract

This research aims to compare the predictive quality and processing demand of the classical technique: ordinal logistic regression and machine learning techniques: artificial neural networks and gradient boosting. The study is set in a context where technological progress has allowed exponential growth in the production of information, which needs to be analyzed efficiently, therefore, it is essential to identify the best techniques for analysis. The comparison was made within the framework of the construction of a model that predicts the level of job satisfaction in Ecuadorian householders with a single job. Therefore, the main characteristics of both methodologies were studied and their equivalences in terminology were identified. Subsequently, a quantitative comparison of the predictive quality was made, processing times and peak RAM associated with each of the models built with the three techniques, a resampling process was performed using ten-fold cross validation and 200 models were run per each technique to control the variability of the phenomenon under study. Finally, the level of processing generated was contrasted, taking into account two factors: 1) sample size (real and increased with 37 336 and 373 360 observations, respectively), and 2) number of effective processor cores (one and seven). The results showed that the total prediction error for gradient boosting was 29.5%, concluding that this technique is the most reliable in its predictive task, presenting a high demand for processing, which decreases considerably when working in parallel, that is, when using all processor cores. It is recommended to use gradient boosting in socio-economic studies like the study proposed here.