It's known that during the model training, we hold out the test-set. However, I actually find during deployment, that if to use a new model train on the entire dataset (train+test), actually yield also pretty good results on predicting truly 'unseen' data. I'm look for comments for this approach. Are there any concerns?

In particular, in this approach I would apply transformations of any kind, e.g. standardization/normalization/categorical encoding, over the entire dataset. Isn't it better for deployment? (Compared to traditional approach where all these transformation are only done on train set, which can sometimes fail to encode some categories of data that are absent in train set.)

Any experience sharing and critique of my approach? TIA.


1 Answer 1


Usually the approach is as follows, first find the model that works best :

  1. Divide the data into train & test.(Testing is out of sample)

  2. Do all the preprocessing on training data and apply the same of test using transform data.

  3. Do hyperparam tuning to find the best model

  4. Do evaluation on test data for best model, once satisfied take the model to production

Taking the model to production:

  1. Take the complete train data and use the same preprocessing and best hyperparams found to train the data again but this time on complete data

  2. For classification you may have to adjust the threshold. This is how model's are deployed across industry.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

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