Random forest final stage: Consider the training dataset or the entire dataset? I trained and tuned a random forest classifier using cross validation and train-test split, as in 70% training dataset (which is then cross validated 5 splits, 6 times) and 30% test dataset. With the CV I managed to get the best hyper-parameters and with the test dataset I obtained the performance score. This link suggests that after everything is ready, one should then refit/retrain the model with the same configuration as before but on the entire dataset now (not the training dataset anymore).
However, I wonder if this wouldn't mess with the hyper-parameters that were defined before specifically for the training data. Also, I wonder if this would not somehow change the nature of the model (as it would most likely start getting 100% accuracy in the entire dataset). This is particularly important for me because I need to use the model on another similar dataset to predict synthetic values. How do you guys usually proceed?
 A: The idea of your test set is to help you objectively evaluate your model's performance on unseen data. If you selected your hyperparameters with CV (on the training set) and you have only run the test set once in the end for a final evaluation, you can consider it to depict your model's performance.
If you are OK with this performance, you can retrain your model on the whole data and expect it to perform at least on par to what it did on the test set (assuming that the new unseen data follows the same distribution as your training/test sets).
Your model's accuracy on the (new) training set doesn't tell you anything. The performance you should look at is that which your model had, when it was objectively evaluated.
Likewise you should use the same hyperparameters as the previous step because they were the ones that achieved the performance you wanted. Changing them after merging the dataset could result in deteriorating the actual model's performance (though you would't know, because now there is no way of objectively evaluating your model). One thing also to note is that the addition of 30% more data shouldn't affect their selection by much.
