Question about PC-Softmax adaptation for multi-label tasks

[ILoveSorasakiHina]

Dear authors,
Thank you for this excellent contribution. Your paper has provided valuable insights for my research. However, I have a question concerning the PC-Softmax method.
My task involves multi-label classification, which essentially consists of multiple independent binary classification problems requiring sigmoid activation rather than softmax. I was wondering if the PC-Softmax approach can be directly adapted to work with sigmoid functions for such scenarios.
I appreciate your time and look forward to your feedback.
Best regards,

[juice500ml]

Yes, it is likely that it will work with minor modifications. Subtracting [log P(y=1), log P(y=2), … log P(y=C)] will be the same. However, for single-label case, P(y=1)+P(y=2), …, P(y=C)=1, which is not true for multilabel case. Also, you’ll have to use sigmoid, not softmax.