1
$\begingroup$

Suppose that my task is to label news articles; that is, to classify which category a news article belongs to. Using the labelled data (with old labels) that I have, I have trained a model for this.

For relevancy purposes, certain labels may be split into multiple new labels. For example, 'Sports' may split into 'Sports' and 'E-Sports'. Because of these new labels, I will need to retrain my model. However, my training data is labelled with the old labels. What can I do to address these 'label updates'?

My idea: Perhaps use some unsupervised clustering method (K-means?) to split the data with the old labels into the new labels. (But how can we be certain that which cluster has what new label?) Then use this 'updated' data to train a model. Is this correct?

$\endgroup$
  • $\begingroup$ so you want to train your new model without having the new data? My first guess would be to relabel the old data. $\endgroup$ – Lustwelpintje Oct 31 at 14:27
  • $\begingroup$ Yes, I want to train my new model without having the new data. Manually relabelling the old data seems impractical to me if my dataset is large, hence I thought of using K-means to do the splitting. I am unsure if this is correct. $\endgroup$ – Air Christmas Nov 1 at 16:16
1
$\begingroup$

Doing some unsupervised learning will give you some divisions within each label but you can't be sure it will be split on "E-sports" and "Sports". It might as wel split sports into "land sports" and "water sports" or "long sports" and "short sports".

The only way to reliably split "sports" in 2 subsections of your desire is to relabel this manualy I am afraid.

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.