- The paper introduces an optimization-based matchmaking framework that maximizes player engagement via predictive churn models and graph matching.
- It employs maximum-weight matching to dynamically arrange players, surpassing conventional skill-based methods and enhancing retention by 0.3%–5.5%.
- The framework integrates adaptive engagement metrics and accommodates multi-player, variable team sizes, offering practical benefits for online platforms.
Summary of EOMM: An Engagement Optimized Matchmaking Framework
Introduction
The paper "EOMM: An Engagement Optimized Matchmaking Framework" (1702.06820) introduces a principled framework for matchmaking in online competitive games, departing from conventional models centered on skill-balancing. The authors formalize matchmaking as an optimization problem with the explicit objective of maximizing player engagement, and propose a computational architecture that aligns match arrangement protocols with quantifiable engagement metrics. This work challenges the longstanding assumption that skill-balanced matchmaking is inherently optimal for player retention and offers a novel algorithmic basis for system design in digital games.
Methodological Framework
The EOMM model is predicated on two primary components: (i) a predictive engagement model that quantifies a player's probability of churning based on match outcomes and context, and (ii) an optimization algorithm that determines matchup pairs (or teams) such that the aggregate engagement — typically inverse churn risk — across all participants is maximized.
Formally, players are nodes in a graph, with edge weights corresponding to projected engagement deltas resulting from potential match assignments. The matchmaking task reduces to identifying edge assignments that globally optimize the engagement-based objective, operationalized via maximum-weight matching algorithms for pairs or graph partitioning for teams. The framework is agnostic to the underlying engagement model, which can be fit using behavioral, demographic, or outcome-driven features.
Empirical Evaluation
The paper provides extensive large-scale evaluations using real-world data from an operational team-based online game. Empirical results demonstrate that EOMM consistently surpasses baseline skill-based matchmaking schemes (including Elo and TrueSkill-based systems) in enhancing short-term player retention metrics. Notably, the authors report statistically significant increases in aggregate engagement rates, with improvements over skill-based methods ranging from 0.3% to 5.5% depending on player population and engagement forecasts.
Furthermore, the framework accommodates dynamic, population-level constraints and exposes trade-offs between fairness, challenge, and engagement optimization. Analysis reveals that, contrary to prevailing industry practices, strict skill balancing is only optimal under restrictive and unrealistic assumptions (e.g., identical engagement responses to match outcomes). The framework generalizes to multi-player and variable team-size settings, showing robust performance even under non-uniform dynamic churn models.
Theoretical and Practical Implications
The theoretical contribution lies in the formalization of matchmaking as a combinatorial optimization problem with engagement-centric objectives, as opposed to classical competence-centric paradigms. This reframing introduces new algorithmic challenges and opportunities; for instance, it enables adaptive matchmaking architectures that personalize player experiences for maximal retention. Practically, EOMM facilitates system-level improvements in live game services, directly impacting metrics crucial for the sustainability of online platforms (DAU, ARPU, LTV).
The paradigm shift suggested by EOMM has implications for the design and operation of digital economies; engagement-optimized matchmaking can drive long-term ecosystem stability, reduce churn rates, and offer differentiated experiences. The framework’s modular integration with existing engagement prediction models accommodates continual learning from behavioral data streams.
Future Directions
The EOMM approach opens avenues for further research, including:
- Integration with reinforcement learning to refine engagement models and dynamically adapt matchmaking protocols in non-stationary environments.
- Exploration of multi-objective optimization incorporating business metrics (e.g., monetization) and social factors (e.g., toxicity).
- Extension to multi-session retention modeling and network-level impacts (e.g., community formation).
- Empirical studies on psychological effects and fairness perception arising from engagement-centric matchmaking.
Conclusion
"EOMM: An Engagement Optimized Matchmaking Framework" (1702.06820) provides a rigorous and scalable architecture for matchmaking in online competitive games, demonstrably outperforming conventional skill-based systems in driving player engagement. The formalization of engagement as the optimization objective establishes a principled foundation for future matchmaking research and system design, with broad implications for both the theory and practice of user retention in interactive platforms.