Regarding con of classifier
How can we justify probability estimates are not to be completely trusted ?
Thank you for your question!
In the lesson, we make the difference between "predictions" and "prediction estimates". The former refers to the predictions themselves (a message is classified as a spam, or a ham). The latter refers to the certainty with which a prediction is made (a message is classified as spam with 70% certainty). One of the lessons that come later in the course ("The YouTube Dataset: Changing the priors") discusses why these probability estimates should not be entirely trusted.