Now y is discrete: y ∈ {0,1} for binary classification.

Why not just use linear regression on 0/1 labels? Squared error isn't the right loss for this — it doesn't map cleanly to a probability or a clean decision boundary. We need purpose-built methods.

Three different philosophies show up in this session and the next: instance-based (KNN), rule-based (decision trees), and generative vs. discriminative — modeling how data looks per class, vs. modeling the boundary directly.