Using Machine Learning to Predict Game Outcomes Based on Player-Champion Experience in League of Legends

Key insights

6 takeaways

01
Machine learning models can predict League of Legends match outcomes with approximately 75% accuracy based solely on pre-game player-champion experience.See it on page 5
02
Gradient Boosting and Deep Neural Networks were the most effective models, achieving 75.4% and 75.1% predictive accuracy respectively using 2020 North American ranked match data.See it on page 4
03
Champion mastery points and specific player-champion win rates are the primary statistical drivers of match success, while total games played and recent activity levels are insignificant.See it on page 2
04
Current Elo-based matchmaking systems fail to account for individual champion proficiency, leading to notable, predictable variances in team skill levels.See it on page 4
05
Players can statistically increase their win probability by focusing on mastering a limited set of champions rather than maintaining a broad, less-specialized pool.See it on page 4
06
The study utilized data from 5,000 unique ranked matches extracted via the Riot Games API to validate the impact of mechanical expertise on competitive outcomes.See it on page 2

Summary

This research investigates the predictability of competitive League of Legends match outcomes by analyzing player-champion experience prior to the start of gameplay. The study aims to determine if mechanical expertise—defined as a player's proficiency with a specific character—serves as a significant predictor of victory, independent of the game's existing skill-based matchmaking system.

The methodology involved collecting data from 5,000 unique ranked matches on the North American server during 2020. Researchers utilized the public Riot Games API to extract features including champion mastery points and player-champion win rates. Through Pearson’s correlation testing, the study identified these two metrics as the primary drivers of match outcomes, while total games played and recent activity levels were found to be statistically insignificant. Several machine learning models were trained and evaluated, including Support Vector Classifiers, k-Nearest Neighbors, Random Forest, Gradient Boosting, and Deep Neural Networks.

The findings demonstrate that match outcomes can be predicted with approximately 75% accuracy before a game begins. The Deep Neural Network achieved a 75.1% accuracy rate, while Gradient Boosting reached 75.4%. These results suggest that despite the implementation of Elo-based matchmaking, there remains a notable variance in team skill levels due to individual champion proficiency. The study concludes that individual mechanical expertise is a critical determinant of success, implying that players can improve their win probability by mastering a limited set of champions. The authors suggest that these findings highlight a potential limitation in current matchmaking algorithms, which do not account for champion-specific experience, and recommend that future research explore player-role experience and broader regional demographics.