Predicting Quality of Video Gaming Experience Using Global-Scale Telemetry Data and Federated Learning

Predicting video game performance through hardware requirements often fails to account for the complex interplay between diverse user configurations and software optimization. To address this, a federated learning framework utilizes global-scale telemetry data from 100,000 users across 835 games to accurately forecast frame rate distributions. By employing learnable knowledge kernels, the system captures both individual player habits and game-specific nuances, effectively mitigating cold-start issues for new player-game pairings while maintaining strict user privacy standards.

Hardware specifications, particularly GPU performance, CPU family, and RAM, serve as the primary determinants of the 95% FPS floor, exerting a significantly greater influence on the gaming experience than game-specific characteristics. Beyond technical hardware metrics, macro-level socioeconomic factors—specifically a nation’s GDP per capita and Gini index—explain 62.5% of the variance in average gaming performance. This indicates that regional economic development is a critical, often overlooked variable in predicting the quality of experience for global player bases.

The federated learning model achieves a mean Wasserstein distance of 0.469, demonstrating that decentralized training maintains high predictive accuracy comparable to centralized approaches. This methodology provides a scalable, lightweight solution for developers and platforms to manage performance expectations and reduce commercial risks associated with poor optimization. By bridging the gap between technical telemetry and user-centric quality of experience, this framework offers a transparent, privacy-compliant mechanism for assessing gaming performance across heterogeneous global hardware environments.