Is it a good idea to retrain ML model on new observations with labels predicted by model itself?

Question

in the company where I work we retrain ML models regularly every day. Now we started to experiment with retraining a model by new observations with labels predicted by model itself.

I've tried to search on the Internet if someone else uses this technique but I didn't find much.

My gut feeling tells me that it is not the right approach as the labels do not represent ground truth any more and therefore the model might deviate and learn an incorrect hypothesis in time.

Am I right that this approach is not correct? Could you please direct me to some sources where I can read more about this topic?

Thank you,

Tomas

Answer 1

You're right, there is an obvious bias: this model will (partly) learn from its own mistakes, totally unaware that these are errors. Thus cases which are rare and/or difficult for the model are likely going to be predicted more and more wrongly, progressively decreasing performance on this kind of cases.

Note that as long as a true gold-standard is used for evaluation, this decrease in performance would be detectable. However if predicted labels are also used for evaluation then this could give the false impression that the model is getting better and better! Needless to say, it would be a very serious mistake to evaluate on predicted labels.

Answer 2

I think you have sort of answered yourself, predictions shouldnt become training data in this way.

Tangentally this sort of idea can work when you have a secondary model that uses the additional information (along with the original ground truth).Look at Psuedo-labelling or surrogate labelling.