I. Pen-and-paper [12v]
Given the following decision tree learnt from 20 observation using Shannon entropy, with leaf annotations (#correct/#total)
- [4v] Draw the training confusion matrix.
- [3v] Identify the training F1 after a post-pruning of the given tree under a maximum depth of 1.
- [2v] Identify two different reasons as to why the left tree path was not further decomposed.
- [3v] Compute the information gain of variable y1.
II. Programming [8v]
Considering the pd_speech.arff dataset available at the homework tab:
- [6v] Using sklearn, apply a stratified 70-30 training-testing split with a fixed seed
(random_state=1), and assess in a single plot the training and testing accuracies of a decision tree with no depth limits (and remaining default behavior) for a varying number of selected features in {5,10,40,100,250,700}. Feature selection should be performed before decision tree learning considering the discriminative power of the input variables according to mutual information criterion (mutual_info_classif).
- [2v] Why training accuracy is persistently 1? Critically analyze the gathered results.
