Probably Approximately Correct Labeling via Multi-Source Allocation

The Journal of Geometric Analysis · Springer Nature · 2026-10-28 · DOI: 10.1007/s11424-026-6205-5

Аннотация

The authors study dataset labeling with inexp ensive but unreliable sources, such as AI models or non-expert annotators, aiming to produce labels whose error rate is controlled with high probability. The authors propose MSPAC, a multi-source labeling procedure that allocates error and confidence budgets across sources and learns an uncertainty threshold for each source. Each source labels only samples with uncertainty below its threshold, and samples that no source considers reliable are sent to experts. The authors formulate allocation as a cost-reduction problem relative to full expert labeling and theoretically show that MSPAC attains this error guarantee and is asymptotically cost-efficient. Comprehensive experiments confirm that MSPAC is empirically valid and achieves superior cost reduction compared with existing approaches.

Читать у издателя →

Похожие статьи