League of Legends: Real-Time Result Prediction

1. Problem Statement

The paper addresses the challenge of predicting match outcomes in League of Legends (LoL), a complex Multiplayer Online Battle Arena (MOBA) game, in real-time. Unlike static pre-match predictions, real-time prediction requires dynamic analysis of evolving game states. The difficulty lies in the non-linear progression of the game, where different variables (e.g., gold, kills, objectives) hold varying levels of importance depending on the match stage (early vs. late game). The study aims to identify the most effective machine learning models and feature sets to predict the winner with high accuracy at specific intervals of elapsed match time, catering to both strategic player improvement and the burgeoning esports betting industry.

2. Methodology

A. Data Collection and Preprocessing

• Source: Data was harvested from the official Riot Games API using the riot-watcher Python library.
• Dataset: The study utilized 64,556 ranked matches played between January 12, 2023, and May 18, 2023, covering skill levels from Iron to Diamond.
• Preprocessing:
  • Time values were converted from milliseconds to minutes.
  • Matches shorter than 5 minutes were excluded to remove remakes.
  • Boolean variables were converted to binary (0/1).
  • Redundant variables (e.g., redWin when blueWin is available) were removed.
• Temporal Segmentation: The dataset was sliced into four distinct scenarios based on Percent Elapsed Time (PET): 20%, 40%, 60%, and 80% of the total match duration. This allowed the models to be evaluated at different stages of the game.

B. Feature Engineering

• Variables: The model utilized 36 features representing both teams (Blue and Red), including:
  • Combat: Champion kills, first blood.
  • Objectives: Dragons (various types), Rift Herald, Baron Nashor, Towers, Inhibitors.
  • Economy/Stats: Total gold, minions/jungle monsters killed, average player level.
• Correlation Analysis: Pearson correlation coefficients were calculated to determine the relationship between features and the Blue team's victory. The analysis revealed that feature importance shifts over time (e.g., early game relies heavily on First Blood and Gold, while late game relies on Dragon/Tower control).

C. Machine Learning Models
The authors evaluated a diverse set of algorithms:

• Logistic Regression
• Random Forest (RF)
• Naive Bayes
• Gradient Boosting (scikit-learn)
• XGBoost
• LightGBM
• Neural Networks (NN)
• Bagging

D. Evaluation Strategy

• Technique: Cross-validation was used to ensure robustness.
• Input: The percentage of elapsed time was included as an explicit input attribute to help models learn temporal patterns.
• Hyperparameter Tuning: Random Search was applied to optimize the hyperparameters for LightGBM and Random Forest.
• Metrics: Accuracy, Precision, Recall, F1-score, and AUC-ROC.

3. Key Contributions

1. Temporal Dynamics Analysis: The study explicitly demonstrates that feature importance is not static. For instance, redTotalGold is the most critical predictor at 20% PET, whereas blueChampionKill becomes dominant at 80% PET.
2. Real-Time Prediction Framework: By slicing data into PET intervals, the paper provides a framework for dynamic prediction rather than a single static outcome.
3. Model Benchmarking: A comprehensive comparison of eight different ML algorithms across four distinct game stages, identifying LightGBM as the superior performer for mid-to-late game scenarios.
4. Unpublished Data Utilization: The use of a fresh, large-scale dataset (2023) ensures the findings reflect the current "meta" of the game, unlike older studies.

4. Results

A. Model Performance by Time Interval

• Early Game (20% PET): Logistic Regression performed best (Accuracy: ~62.28%), followed by Gradient Boosting and XGBoost. This suggests linear models capture early-game economic disparities well.
• Mid-to-Late Game (40% - 80% PET): LightGBM consistently outperformed all other models.
  • At 60% PET, LightGBM achieved 77.88% accuracy.
  • At 80% PET, LightGBM achieved 85.42% accuracy (the highest overall).
• Optimization: Applying Random Search to LightGBM and Random Forest yielded marginal but consistent improvements in accuracy (e.g., LightGBM at 80% PET improved from 85.42% to 85.48%).

B. Feature Importance Trends (Random Forest Analysis)

• 20% PET: redTotalGold was the most important feature.
• 40% & 60% PET: blueTotalGold became the dominant predictor.
• 80% PET: blueChampionKill emerged as the most significant factor, indicating that team fight dominance becomes the primary driver of victory in the late game.

C. Correlation vs. Feature Importance
The study noted a divergence between simple correlation and model-based feature importance. While correlations show linear relationships, feature importance (via Random Forest) accounts for non-linear interactions and variable dependencies, providing a more accurate reflection of what drives the model's decision-making.

5. Significance and Conclusion

• Practical Application: The results are highly relevant for the esports betting industry, offering tools for more accurate in-play wagering. They also provide players and coaches with data-driven insights into which objectives are most critical at specific game stages.
• Scientific Contribution: The paper validates that incorporating "time elapsed" as a feature significantly boosts prediction accuracy. It highlights that a "one-size-fits-all" model is suboptimal; the best model depends on the game state.
• Future Work: The authors suggest future research should focus on real-time latency and processing rates to ensure these models can be deployed live during matches without lag, as well as exploring the integration of player-specific behavioral data.

In summary, the paper establishes that LightGBM is the state-of-the-art approach for LoL match prediction, particularly when the match is past the 40% mark, achieving an accuracy of over 85% in late-game scenarios.