Comfort pick or meta pick? Predicting League of Legends matches from champion experience
- Article
- Games
League of Legends (LoL) is a multiplayer online battle arena game. In ranked play, two teams of five each pick a champion and try to destroy the other team’s base. Fair matchmaking matters for how the game feels, and the game handles it with an Elo rating system like the one used in chess. What the matchmaking does not consider is how much experience a player has with the specific champion they picked, even though that plays a large part in who wins. Tiffany D. Do, Seong Ioi Wang, Dylan S. Yu, Matthew G. McMillian, and Ryan P. McMahan built a model to predict the outcome of a ranked match from that experience alone, ignoring team composition.
Approach
The authors used ranked matches from the North American server in 2020. A Pearson correlation test picked the features that actually relate to the outcome: player-champion win rate and champion mastery points were correlated with the result, while total games played and the number of recent games on the chosen champion were not.
They compared five models chosen for their record on MOBA outcome prediction: support vector classifiers (SVC), k-nearest neighbors (kNN), random forests (RF), gradient boosting (GBOOST), and deep neural networks (DNN). The data was split into 4,000 training and 1,000 test matches, and evaluated with stratified 10-fold cross-validation, which preserves the class balance in every fold.
Findings
Every model predicted the winner with better than 70 percent accuracy. Gradient boosting scored highest at 75.4 percent, with the DNN close behind at 75.1 percent. Because the GBOOST model also had the largest standard deviation (5.25 percent) and standard error of any model, the authors chose the DNN, at 1.9 percent standard deviation, as the most appropriate. Since the data came from one server, they note that other regions with different player bases could give different results and suggest repeating the study elsewhere.
What it means for players
With 162 champions, no two ranked games look alike, and the matchmaking system adds to that variance by ignoring player-champion experience. Most players only play a handful of champions regularly. The result here says that a match can be called with 75 percent accuracy from champion experience before the game even starts, which is a good argument for the comfort pick over the meta pick.
Source: Do, T. D., Wang, S. I., Yu, D. S., McMillian, M. G., and McMahan, R. P. (2021). Using Machine Learning to Predict Game Outcomes Based on Player-Champion Experience in League of Legends. Proceedings of the 16th International Conference on the Foundations of Digital Games. doi:10.1145/3472538.3472579, also on arXiv:2108.02799.