Approximating the Core via Iterative Coalition Sampling

The core concept offers practical value for AI explainability: The paper demonstrates that the core from cooperative game theory can complement existing methods like Shapley values to identify which features or data instances are most influential in black-box AI model predictions, providing a complementary approach to understanding how AI systems make decisions.

Computational barriers to the core are being overcome: Historically, calculating the core has been computationally prohibitive because it required solving large linear programs. This paper proposes new iterative algorithms that avoid this bottleneck, making core-based analysis feasible for much larger problems than previously possible.

The new approach scales to real-world problem sizes: The authors demonstrate their algorithms work across different types of cooperative games (weighted voting games, induced subgraph games, and marginal contribution networks) and scale better than traditional methods, suggesting practical applicability to real AI systems.

The core identifies when coalitions have incentive to "defect": By finding allocations where no subset of agents wants to break away and form their own group, the core provides insights into stable solutions—relevant for understanding fairness and stability in AI-driven decision-making systems.

This work bridges game theory and AI transparency: The paper connects cooperative game theory concepts to explainable AI (XAI), suggesting that mathematical frameworks from economics can help make AI systems more interpretable and trustworthy to stakeholders.