Role-Sensitive Analysis of Positional Contributions and Win Prediction in Multiplayer Online Battle Arena Esports

Abstract

This study developed a role-sensitive framework for evaluating player contributions and predicting match outcomes in League of Legends, a leading multiplayer online battle arena (MOBA) esports title. Unlike prior approaches that treat the team as a homogeneous unit, we introduced the Victory Contribution (VC) metric to quantify individual player impact across four strategic domains—offense, defense, laning, and gold acquisition—while accounting for positional roles. Using professional match data from 10,390 games across three global leagues in 2023 and an independent test set of 9,670 matches from 2024, we trained four classification models—logistic regression (LR), multilayer perceptron (MLP), support vector machine (SVM), and XGBoost (XGB)—on 13 selected features, both with and without the VC metric. Across all classifiers, extended models incorporating VC consistently outperformed baseline models. On the test set, the XGB model achieved the highest performance with an accuracy of 0.8686, recall of 0.8084, F1-score of 0.8601, and AUC of 0.9512. VC also outperformed traditional metrics such as Damage Share (DS) and Kill Participation (KP), yielding higher test recall and F1-score than both. All observed performance improvements were statistically significant across evaluation metrics (p < 0.0001). Cross-league analysis revealed no significant differences in positional VC ratios, suggesting structural consistency in role contributions across Legends Champions Korea (LCK), Legends European Championship (LEC), and Legends Championship Series (LCS). These findings demonstrate the practical value of incorporating role-aware metrics in predictive modeling and player evaluation in professional esports.