Prediction Models Based on Players' Historical Data for Evaluating Competitive Balance in League of Legends Matches (Modelos de Previsão Baseados em Dados Históricos dos Jogadores para Avaliar do Equilíbrio Competitivo em Partidas de League of Legends)

Resumo

https://doi.org/10.5281/zenodo.15702048

This study aims to investigate possible trends in the ranked queues of the game League of Legends, developing predictive models to evaluate the balance of matches based on players' historical data. The data was collected through the game's official API, including statistics such as KDA, gold per minute, vision score, damage per minute, and objectives achieved (towers, dragons, and barons). Different machine learning techniques, including Random Forest, Logistic Regression, and SVM, were implemented and compared, using Python for backend development and data processing.

The models were evaluated using metrics such as accuracy, F1 Score, recall, and support. The Random Forest model showed the best performance during training, achieving 96% accuracy. However, when tested with live match data, its performance dropped significantly, with the best result reaching 48%. The confusion matrix revealed the models' limitations in capturing more complex nuances of the matches. Additionally, an analysis of the win-rate proportion between teams was conducted, contributing to a broader understanding of the factors influencing match balance.

The results suggest that, although the selected attributes are relevant, they are insufficient to identify patterns indicating imbalance in matches. This evidence highlights the effectiveness of the matchmaking system and the impact of Matchmaking Rating (MMR) on maintaining balance in matches.