I'm trying to do a Kaggle competition and I started doing a first model with all the features and in the top 5 features I noticed that one in particular contains many missing values (but still in top 5).
I had the idea to fill the missing values by trying to fit a classification model using the other features (apart from the target) and predict those missing values.
This didn't improve the model that much. There is other features that are in the same case that can be also filled using this technique.
My question here is: Is this a good idea ? if yes how far can I go with this ? Any thoughts on this subject would be welcome.