Notes:
A model trained in the region covered by the red ellipse would try to make predictions based on this knowledge even if the input space lies in the region covered by the blue ellipse. Hence, if we have a wide range of dataset covering the whole region the predictions would be much reliable.