What is Supervised Learning?
Supervised learning is a machine learning approach in which a model is trained on examples that pair an input with the correct output, called a label. During training the model adjusts its parameters to reduce the difference between its predictions and those labels. Once trained, it assigns outputs to new inputs that were never labeled.
Tasks divide into classification, where the output is a category such as fraudulent or legitimate, and regression, where the output is a continuous quantity such as a price or a duration. The same training machinery applies to both, differing mainly in the loss function used to measure error and in the metrics used to report performance to stakeholders.
Data is normally split into training, validation, and test sets. The model fits the training set, hyperparameters are tuned against the validation set, and the test set is reserved for a single honest estimate of performance on unseen data. Reusing the test set for tuning inflates reported accuracy and is a common source of disappointing production results.
The main practical constraint is labels. Annotating data is slow, and labels can be inconsistent, ambiguous, or systematically biased by who produced them and under what instructions. Techniques such as transfer learning, weak supervision, and synthetic data reduce the number of hand labeled examples required, but the label definition itself still determines what the model learns to optimize.
Supervised learning is also used to shape general purpose models. Instruction tuning fits a pretrained model on curated pairs of prompts and preferred responses, which is supervised learning applied on top of self supervised pretraining. Class imbalance, label noise, and distribution shift after deployment are the recurring failure modes across all of these settings.
Key points
- Trains on pairs of inputs and correct outputs.
- Covers classification and regression tasks.
- Requires separate training, validation, and test splits.
- Labeling cost and label bias are the main limits.
- Distribution shift degrades accuracy after deployment.
In practice
A lender builds a model to estimate default risk. Historical applications supply the inputs, and whether each loan was repaid supplies the label. The model learns the association and scores new applicants. If historical decisions were biased, the labels encode that bias and the model reproduces it, which is why label provenance is audited alongside accuracy.