What supervised learning is
Supervised learning is the most widely used paradigm in machine learning. You give the model many examples where each input is paired with the correct output, called a label. The model studies these pairs and gradually learns a mapping from inputs to outputs, so that when you later show it a brand-new input, it can predict the right label. The word “supervised” refers to this guidance: every training example acts like a teacher providing the right answer.
Features, labels and predictions
Two pieces of vocabulary describe the data:
- Features — the measurable inputs describing each example, such as the pixels of an image, the words in an email, or the size and location of a house.
- Label — the correct answer attached to that example, such as “cat,” “spam,” or the sale price.
During training the model produces a prediction, compares it to the true label, measures the error with a loss function, and adjusts itself to reduce that error. Over many examples the predictions get closer to the labels.
Classification vs regression
Supervised problems split into two families based on what the label looks like:
- Classification predicts a category from a fixed set. Examples: detecting spam vs not-spam, diagnosing whether a scan shows a tumour, or identifying which of ten digits a handwritten number is.
- Regression predicts a continuous number. Examples: forecasting tomorrow’s temperature, estimating a house price, or predicting how many units a product will sell.
The underlying idea is identical — learn from labelled pairs — only the type of output differs.
Real-world examples
Supervised learning powers a huge amount of everyday technology:
- Email spam filters trained on messages humans marked as spam or legit.
- Credit scoring that predicts default risk from past borrowers labelled as repaid or defaulted.
- Medical imaging models trained on scans labelled by radiologists.
- Recommendation ranking that learns from items users clicked or ignored.
Splitting the data
To know whether a model genuinely learned (rather than memorised), the labelled data is split into three parts:
- Training set — the examples the model learns from.
- Validation set — used to tune choices like model size or learning rate.
- Test set — held back until the very end to estimate real-world accuracy.
If a model does well on training data but poorly on the test set, it has overfit — memorised noise instead of learning the underlying pattern. Good supervised learning is all about generalising to data the model has never seen.