TLDR: This research introduces an explainable machine learning system that predicts e-sports match outcomes in real-time, achieving over 90% accuracy. It uses dynamic data management with sliding windows to adapt to game changes and provides clear explanations for its predictions through natural language and a visual dashboard, fostering trust and aiding player decision-making.
The world of electronic sports, or e-sports, has exploded in popularity, drawing millions of players and spectators globally. This rapid growth has also fueled the demand for advanced analytics, particularly in predicting match outcomes. While Artificial Intelligence (AI) has traditionally been used to find patterns in game data, most efforts in predicting wins have focused on analyzing data in large batches, often overlooking the need for real-time insights and clear explanations of how predictions are made.
A new research paper, titled “Explainable e-sports win prediction through Machine Learning classification in streaming,” addresses these critical gaps. Authored by Silvia GarcÃa-Méndez and Francisco de Arriba-Pérez from the Information Technologies Group at atlanTTic, University of Vigo, this work introduces a groundbreaking solution for predicting e-sports wins in real-time, with a strong emphasis on making these predictions understandable to users. You can read the full paper here.
The Challenge: Dynamic Games and Black Box AI
One of the biggest hurdles in e-sports prediction is the dynamic nature of games. Unlike traditional sports, e-sports like Counter-Strike: Global Offensive (CS:GO), Dota 2, or League of Legends constantly evolve with updates that can significantly alter game mechanics. This means that historical data can quickly become outdated. Furthermore, many AI models operate as “black boxes,” meaning their internal workings are opaque, making it difficult for players, coaches, or commentators to understand why a particular prediction was made. This lack of transparency can erode trust in the AI’s recommendations.
A Real-time, Explainable Solution
The researchers developed an innovative Machine Learning (ML) system designed to overcome these limitations. Their solution operates in streaming, meaning it processes data as it arrives, rather than waiting for large datasets to accumulate. A key aspect of this real-time operation is the use of “sliding windows,” which allow the system to continuously analyze recent game changes and adapt its predictions accordingly. This is particularly crucial for games like CS:GO, which was the focus of their study, known for its complex round-based mechanics and strategic depth.
How the System Works
The system is built on several interconnected modules:
In-game Data Fusion: This initial step gathers and combines various data points from the game, such as event identifiers, maps played, timestamps, and individual player statistics. It aggregates data for each of the five players per team, creating a comprehensive snapshot of player skills at a team level.
Stream-based Data Processing: To ensure the ML models perform effectively in a streaming environment, this module engineers new features from historical player data using the sliding window technique. It calculates statistics like averages, percentiles, and standard deviations over different timeframes. It also includes a feature analysis and selection step, which intelligently discards less relevant data to maintain efficiency.
Stream-based Classification: This is where the actual win prediction happens. Unlike traditional batch learning, the ML models (including Hoeffding Adaptive Tree Classifier and Adaptive Random Forest Classifier) are continuously updated with every incoming game sample. The system first makes a prediction and then retrains the model based on the new data, ensuring it remains current and accurate. The Adaptive Random Forest Classifier, despite its higher computational cost, proved to be the most accurate, processing 22.5 samples per second and achieving high performance.
Post-game Stream-based Explainability: This is the core innovation for user trust. After a prediction, the system provides clear explanations. It uses post-game features and traverses the decision paths of the tree-based ML models to identify the most influential factors in a prediction. These explanations are presented in natural language descriptions and a visual dashboard, helping players understand their performance, anticipate rewards or penalties, and identify areas for improvement.
Also Read:
- AI Predicts VALORANT Round Outcomes Using Tactical Video Analysis
- Bridging LLM Flexibility and Rule-Based Reliability for RTS Games
Impressive Results and Future Outlook
The experimental results were highly promising, with the system achieving an accuracy higher than 90% in win prediction, significantly outperforming existing solutions in the literature. This high accuracy, combined with the explainability module, makes the system a powerful tool for ranking and recommender systems, fostering greater trust in AI-driven insights.
Looking ahead, the researchers plan to integrate live API information into their model and explore how new features, such as players’ physiological data, might influence win predictions. They also aim to make the system available to the research community to gather user feedback on the explainability dashboard and explore other classification tasks like death prediction.
This research marks a significant step forward in e-sports analytics, offering a robust, real-time, and transparent approach to win prediction that can empower players and enhance the overall e-sports experience.
