Applying Machine Learning to Predict League of Legends Match Outcomes Using Draft-Phase Data

Abstract

This study investigates the effectiveness of Autoencoder-based Champion Embedding for predicting League of Legends match outcomes during the Draft Phase. Utilizing a dataset of 5,612 professional matches from the 2020 season, the research aims to address the complexity and noise of high-dimensional champion features that often degrade model generalization. The methodology involved compressing 28 initial champion attributes into low-dimensional vectors ranging from 2D to 16D, with strict data leakage prevention protocols where feature standardizers were fitted exclusively on the training set. Based on 10 repeated experimental trials (10 Random Seeds) to ensure statistical reliability, 6D Embedding was identified as the "sweet spot," providing the optimal balance between reconstruction accuracy and predictive generalization while minimizing underfitting and overfitting risks. Results demonstrate that 6D Embedding paired with a Soft Voting ensemble model achieved a mean Area Under the ROC Curve (AUC) of 0.5421 ± 0.0161 and a mean Accuracy of 0.5329 ± 0.0123, consistently outperforming established baselines. While performance was numerically close to the high-dimensional Raw Data (AUC 0.5523), a paired t-test yielded a p-value of 0.509, confirming no statistically significant difference at the 0.05 level. This confirms the primary advantage of the proposed framework: reducing feature dimensionality by 81.25% (from 320 to 60 dimensions) while maintaining predictive integrity and reducing data noise. These findings suggest that Champion Embedding is a promising approach for enhancing strategic analytics and decision-support systems in the professional esports industry.