Real-Time Player Engagement Measurement in Video Games

Abstract

Player engagement is crucial for video games, directly impacting satisfaction, retention, and commercial success. Game developers currently rely on post-hoc analytics or sales metrics that cannot capture real-time engagement fluctuations, while research approaches depend on intrusive methods requiring specialized equipment, creating a gap between practical needs and current capabilities.

This thesis investigates non-intrusive player engagement measurement methods for both game developers seeking practical optimization tools and researchers studying engagement dynamics. It identifies Flow Theory - which posits optimal engagement occurs when skill and challenge are balanced - as a promising framework for real-time prediction.

The MultiPENG study evaluated engagement across multiple modalities, revealing human judges achieved only 50% accuracy with poor inter-rater agreement (Krippendorff's α = 0.04-0.09). Computational approaches demonstrated effective performance, with facial footage (63% accuracy), EEG signals (61%), and eye metrics (59%) showing that the webcam-based approach offered the best balance between performance and practicality. Most significantly, a model using only player skill and challenge as predictors (67% accuracy) performed on par with complex multimodal approaches (65% accuracy), empirically validating Flow Theory. Despite these promising results, the observed "cold-start" sensitivity suggests careful interpretation when generalizing to new participants.

Building on these insights, a novel telemetry-based framework was developed using PlayerUnknown's Battlegrounds - a challenging case study selected for its complex environment combining shooting, combat, scavenging, survival mechanics, and large-scale multiplayer interactions across diverse gameplay phases. The framework's hybrid architecture combining Graph Convolutional Networks with Transformers outperformed Transformer-only models (73% vs. 67% accuracy). Requiring just one minute of gameplay data, the system can proactively forecast engagement by estimating skill and challenge in future timesteps, and then mapping them to an engagement level. Performance matches questionnaire-based methods while operating non-intrusively with standard game telemetry.

The primary contributions include: (1) the MultiPENG dataset with synchronized data across modalities enabling direct comparison; (2) empirical validation of Flow Theory, demonstrating skill-challenge metrics can match complex multimodal approaches; (3) a methodology for measuring skill and challenge directly from gameplay telemetry; and (4) a real-time engagement framework combining Graph Convolutional Networks with Transformers that enables adaptive experiences without specialized equipment. These contributions serve game developers seeking player experience optimization tools and researchers investigating engagement dynamics in interactive systems.