Low performance of trained classfier when the input labels are clustered

[facebookresearch]

Low performance of trained classfier when the input labels are clustered

[riyadparvez]

Wouldn't it be easier to just shuffle training data before training in fastText itself so that users do not get surprised?

[adam2326]

@riyadparvez No since fastText does not accept IDs separately it would be difficult to join data back if it was part of a larger data set. I dont want to append an ID to the text itself. So shuffle-> break text off -> fasttext -> join. I can count on order this way.

[EdouardGrave]

Yes, this is an expected behavior. FastText uses stochastic gradient descent to learn the model, and the examples are processed "in order". Thus, it is important to shuffle the training data, so that not all the examples of a given class are seen first, then the examples of a second class, etc.

[xiamx]

Thanks for the clarification 👍