Tag
This paper resolves the minimax rate for agnostic multiclass transductive learning, demonstrating that the optimal excess error is governed by the DS and Natarajan dimensions. The result holds for arbitrary label spaces, extending prior work on binary classification.