#decision-tree

A depth-2 decision tree for SMS spam uses binary word features (1 = the word occurs). The training counts in each leaf are given as [ham, spam]:

• call = 0 and txt = 0 → [3649, 229]
• call = 0 and txt = 1 → [10, 120]
• call = 1 and me = 0 → [108, 262]
• call = 1 and me = 1 → [76, 3]

The message "Call me" reaches the leaf for call = 1, me = 1, so the tree predicts "ham". If the tree reports its certainty as the fraction of training examples in the leaf that share the predicted class, how certain is it?