Humans are quite good at pattern recognition, but in some cases, like cancer cell identification or high frequency trading, machines can pick up on subtle patterns invisible to most humans. We hypothesize that this applies to win ratio predictions in MOBA games, specifically League of Legends. League of Legends is a competitive multiplayer game where teams of 5 players choose unique champions and compete to destroy the enemy nexus.
Existing work in this field revolves around computing win probability over the course of the match based on real time data (gold leads, objectives, etc.) [1] [2]. However, being able to predict game outcomes based solely on pregame data, such as player and champion trends, can inform players of how to approach the match to maximize chances of winning. In addition, the ability to predict game outcomes can be used to analyze the fairness of matchmaking and inform better matchmaking algorithms.
The primary dataset will be the Riot games API, which will allow us to request a wide variety of relevant data.
Given only pre-game information available at match start, our goal is to generate a winrate graph over time for a given match. Specifically, we want to model \(\mathbb{P}(\text{win} \mid \text{pre-game features, match duration} \geq t)\)for some time \(t\).
The Riot Games API returns data for player statistics, champion statistics, and match information from different endpoints. Thus the first step in data preprocessing would be to collect the relevant player stats, champion stats, and match stats for every match. The API data will then be augmented with our own variables calculated from API data, like classifing games as one-sided, come-back, etc. Poor quality matches, such as matches that include AFKs or trolling behavior would be culled.
We plan on using various supervised learning methods, such as use a Multi-Layer Perceptron (MLP) with embedding layers to encode Player IDs (to capture personal tendencies), Champion IDs (some are naturally tied to certain roles), Team composition and outputs predicted lanes (Top, Jungle, Mid, Bot, Support) for all five players. In addition, XGBoost can be used to capture nonlinear relationships like champion synergy or role balance.
For qualitative evaluation metrics, we use cross-entropy loss, which penalizes incorrect predictions more heavily when made with high confidence. This encourages probabilistic outputs that reflect true uncertainty, rather than overconfident guesses.
League of Legends is complex and variable, so we do not expect binary win/loss predictions to be always accurate. Instead, we aim for the predicted confidence scores to approximate the true win probability.
Additionally, we expect the predicted win probability curves over time exhibit temporal consistency, meaning they evolve smoothly and respond meaningfully to changes in pre-game features(Runes) and match duration.
The main goal of the project is to create an accurate winrate-over-time graph using pre-game information. This could inform players on how to optimize their strategy for team composition. Additionally, we will analyze the fairness of the matchmaking system, which should ideally be 50% winrate. Ethical concerns could come in the form of using the tool to gamble on games, so certain safegaurds will need to be put into place to prevent this.
We expect to see vastly different winrate graphs based on different team compositions. While player skill in a given match is often unpredictable, factors like champion interactions significantly impact the match result. We expect this model to reinforce some existing intuitions we have about the game, such as seeing “early game” compositions having higher winrates early, that drop off as the game progresses.
[1] J. B. S. Junior and C. E. C. Campelo, “League of Legends: Real‑Time Result Prediction,” arXiv preprint arXiv:2309.02449, Sep. 2, 2023.
[2] D.-H. Kim, C. Lee, and K.-S. Chung, “A Confidence‑Calibrated MOBA Game Winner Predictor,” arXiv preprint arXiv:2006.15521, Jun. 28, 2020.
[3] T. D. Do, S. I. Wang, D. S. Yu, M. G. McMillian, and R. P. McMahan, “Using Machine Learning to Predict Game Outcomes Based on Player-Champion Experience in League of Legends,” Proc. 16th Int. Conf. Foundations of Digital Games, Article 47, pp. 1-5, Aug. 2021.
| Name | Proposal Contributions |
|---|---|
| Nicholas | Drafted motivation and problem statement |
| Yuchen | Designed initial model pipeline and preprocessing plan |
| Owen | Recorded presentation and finished the gantt chart |
| Eric | Created literature review summary and comparison tables |
| Arya | Structured slide content and formatted final presentation |
| Richard | TA |