#base-learners

Question 1: Single choice

mediumAI-checkedDecision Trees & Ensembles

A boosting algorithm with the logistic loss chooses each vote by minimizing $\sum_t \ln\big(1 + e^{-y_t(h_{m-1}(x_t) + \alpha_m h(x_t; \hat\theta_m))}\big)$ over $\alpha_m \ge 0$ . The base learner is a linear SVM, and the training set is linearly separable. What is $\hat\alpha_1$ in the first round ( $h_0 = 0$ )?

It is finite, because the logistic loss flattens out for large margins1
It is 1, the scale at which the SVM's margin constraints are tight2
It is 0, because a perfect base learner needs no reweighting3
It is infinite, since a larger $\alpha_1$ lowers every example's loss4

Timed test