A company labels 100,000 emails as "spam" or "not spam" and trains a binary classifier. Six months later, they build a second model using only email metadata (no labels) to group emails into clusters. Which learning paradigm does each model use, and what structural property of the data drives this distinction?