We target open-world feature extrapolation problem where the feature space of
input data goes through expansion and a model trained on partially observed
features needs to handle new features in test data without further retraining.
The problem is of much significance for dealing with