Title: Encouraging Cooperation in Active Learning

Abstract: This study presents a novel framework for incentivized collaboration within the context of collaborative active learning, a scenario in which several agents seek to acquire labels for a shared hypothesis. The primary objective of these rational agents is to secure necessary labels for their respective datasets while minimizing the associated label complexity. Specifically, the research focuses on the development of (strict) individually rational (IR) collaboration protocols, which are designed to ensure that no agent can lower their expected label complexity by choosing to act independently.

Our analysis demonstrates that when utilizing any optimal active learning algorithm, the straightforward implementation of that algorithm across the entire dataset inherently satisfies the conditions for being individually rational. However, because calculating the optimal algorithm is computationally intractable (NP-hard), we propose alternative collaboration protocols. These proposed protocols maintain (strict) individual rationality and exhibit label complexity performance comparable to that of the most effective known tractable approximation algorithms.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC