Comparación de Cinco Modelos de Machine Learning para la Predicción de las Elecciones Presidenciales en Colombia: una Perspectiva con Datos Composicionales

Abstract

In recent years, numerous studies have employed machine learning techniques and compositional data analysis in various fields of study. However, their integration into electoral analysis remains limited. For this reason, this work combines both approaches by applying five machine learning models: random forest, gradient boosting, support vector machines, k-nearest neighbors, and feedforward neural networks, to predict the results of the presidential elections in Colombia at the municipal level, onsidering the data as compositional. Specifically, it forecasts the vote distribution in each municipality along a unidimensional Left-Right ideological spectrum. This approach aims not only to improve prediction accuracy but also to comtribute a significant advancement in methodologies applied to electoral analysis. The models were trained on 70% of the presidential election data from 2002 to 2022 and evaluated on the remaining 30%. The algorithms demonstrated similar performance across transformations of each ideological spectrum, with variability percentages between 56% and 94% in predicting vote proportions, with the feedforward neural networks model using the centered log-ratio transformation achieving the best results.