Supervised Learning

Supervised learning is the setting where every training example comes with the answer attached. You hand the model pairs — an email and the label spam, a customer record and whether they churned, a photo and the object in it — and training adjusts the model until its predicted answers match the given ones closely enough to generalise to examples it has never seen. It is by far the most commercially deployed form of machine learning, because most business questions are naturally shaped this way: given what we know about this thing, predict the field we care about. The two sub-types are classification, where the answer is a category, and regression, where it is a number. Choosing between them is usually obvious, but the boundary matters for how you evaluate: accuracy is a reasonable measure for balanced classification and a badly misleading one when 98% of your examples belong to one class, in which case a model that always predicts the majority looks excellent and is useless. Precision, recall, and the trade-off between them are what you actually care about when the classes are lopsided, which in business data they usually are. The cost of supervised learning is labels, and this is the practical bottleneck. Someone has to decide the ground truth for thousands of examples, consistently, including the ambiguous ones — and inconsistent labelling puts a hard ceiling on achievable accuracy that no model choice can lift. Teams routinely underestimate this: label quality, not model architecture, is the usual difference between a working classifier and an abandoned one. Where labels already exist as a by-product of the product, supervised learning gets cheap. Support tickets that agents already categorised, emails users already marked as spam, invoices someone already keyed in, subscriptions that already churned or didn't — each is a labelled dataset you did not have to pay for. Spotting that free supervision hiding in your own database is the highest-leverage move available, and it is usually a product-thinking exercise rather than a machine-learning one.