Modelagem preditiva de partidas de league of legends usando aprendizado supervisionado

Abstract

This study explores the application of supervised learning models to predict the outcomes of League of Legends matches, one of the most popular esports worldwide. Various approaches were analyzed, including K-Nearest Neighbors (KNN), Logistic Regression, Gradient Boosting, and Decision Trees, using a dataset extracted from Oracle’s Elixir. The methodology involved careful feature selection, removal of irrelevant metadata, and model validation through metrics such as AUC-ROC and accuracy. Results indicate that Logistic Regression and KNN achieved the best performance, with an AUC of up to 0,930, demonstrating strong predictive capability. Beyond model accuracy, the study highlights the strategic importance of early-game resource differences and champion selection in decision-making within competitive settings.