Cardinality-Aware Set Prediction and Top-k Classification

To optimize this loss function, we propose two families of surrogate losses: cost-sensitive comp-sum losses and cost-sensitive constrained losses.