Unravelling in Collaborative Learning

Collaborative learning, while promising, faces challenges when participants are strategic and data quality varies. A naive approach to aggregation may lead to ‘unravelling’, where only the lowest quality data contributors remain, hindering model performance. This occurs because high-quality contributors withdraw to avoid being negatively affected by lower-quality data.

This paper introduces a novel method that tackles this issue. Using probabilistic verification, the researchers design a mechanism to make the grand coalition (all agents participating) a Nash equilibrium with high probability, thus breaking the unravelling effect. This is achieved without using external transfers, making the solution both effective and practical for real-world deployment. The method is demonstrated in a classification setting, highlighting its applicability to real-world scenarios.