Missing data in train set and test set

Question

I have a dataset of N columns. Now I'm able to preprocess data and find a subset of features that I can use to train a model and make predictions. In the case where the train data has missing feature values, I remove those data points and train on the residua set. In the case where the data I am supposed to predict has missing value for a feature that I am using, how can I predict with my model?

One way I can think of resolving this is by using multiple subsets of features and train separate models to minimize the probability that I will have a feature with "nan" value. However there is still the remote possibility that I will encounter data that will have a missing feature for each model. What techniques can I use here, assuming I have no way of predicting what the missing value might be?

Answer 1

Generally speaking, any missing values you find in the data you want to predict (unseen data or test data) should be handled the same way you handled in the training data. For example, in your training data you decide to impute a missing feature with the mean value, then you would have to use the same value for imputing the missing feature in the test data.