Last reviewed:
What is supervised learning? Definition and difference from unsupervised learning
Supervised learning is a machine learning method in which a model is trained on examples whose correct answer is known (a label), so that it can predict that answer for new cases. It is found in common predictive uses in business: customer scoring, fraud detection, automatic sorting of requests.
The CNIL, the French data protection authority, defines supervised learning as a machine learning process in which the algorithm is trained on a given task using a dataset where each item carries an annotation indicating the expected result. In practice, the model is shown a large number of examples: an email and its category (“complaint”, “quote”, “cancellation”), a transaction marked “fraudulent” or “normal”, an invoice and its accounting code. The model adjusts its parameters to reduce the gap between its predictions and these known answers. Its performance is then measured on labelled examples it has never seen, before it is used on real cases. Two families of tasks dominate: classification, which assigns a category, and regression, which predicts a numerical value (a lead time, an amount, a probability of default). Supervised learning is one of the three main paradigms of machine learning, alongside unsupervised learning, which works on unlabelled data, and reinforcement learning, which learns by trial and reward. Its strength is also its limit: everything depends on the labels. They cost expert time and reproduce the errors or biases of those who assigned them. The European AI regulation takes this into account: for high-risk systems, its Article 10 requires governance of training, validation and test data, including annotation and labelling operations. Large language models also rely on it: in InstructGPT (OpenAI, 2022), supervised fine-tuning on human-written answers precedes the RLHF stage.
Concrete example
Illustrative cases (fictitious companies). A French regional mutual insurer with 300 employees receives several thousand emails from members every month. It has two years of history already categorised by its advisers (reimbursement, membership, cancellation, complaint). It trains a classification model that suggests a category and a processing queue for each new message. Advisers correct the errors, and these corrections become new labels for the next training round. In a different area, a food wholesaler uses its history of unpaid invoices to train a model that estimates the risk of late payment for each business customer. The credit department keeps the decision: the score is used to prioritise reminders and adjust payment terms.
Comparison
| Criterion | Supervised | Unsupervised | Reinforcement |
|---|---|---|---|
| Starting data | Examples labelled with the right answer | Raw data, no expected answer | No fixed dataset: trials in an environment |
| Question addressed | “Which category, which value?” | “Which groups, which anomalies?” | “Which action gives the best result?” |
| Business examples | Scoring, fraud detection, ticket classification | Customer segmentation, anomaly detection, grouping of customer comments | Logistics optimisation, online advertising, tuning AI assistants (RLHF) |
| Main cost | Building and checking the labels | Interpreting results with business teams | Building a reliable environment or reward |
| Main risk | Reproducing the errors and biases of the labels | Finding groups with no business meaning | Optimising the reward at the expense of the real intent |
FAQ
What is supervised learning in simple terms?
It is learning from corrected examples. The model is given past cases together with the right answer (this ticket was a complaint, this transaction was a fraud). It derives statistical rules from them and then applies them to new cases.
What is the difference between supervised and unsupervised learning?
Supervised learning starts from labelled data and learns to predict an answer known in advance. Unsupervised learning starts from raw data, with no expected answer, and looks for groupings or anomalies. The first answers “which category?”, the second “which groups emerge?”.
What are examples of supervised learning in business?
Classifying incoming tickets or emails, detecting payment fraud, scoring customer risk, forecasting sales or delivery times, reading invoices automatically, image-based quality control on a production line. In each case, an annotated history exists or can be built.
Do you need a lot of labelled data?
It depends on how hard the task is and how many categories there are. A simple task with a few clearly distinct categories needs fewer examples than a fine-grained task with rare cases. For text, you can also start from an already trained language model and adjust it on a small volume of in-house examples (fine-tuning), or first test a simple instruction with a few examples.
Does ChatGPT use supervised learning?
Partly. Language models are first pre-trained on large volumes of raw text to predict the next word. Then, as OpenAI's InstructGPT paper (2022) describes, they are fine-tuned in a supervised way on answers written by humans, before a stage of reinforcement learning from human feedback (RLHF).
Is supervised learning regulated by the AI Act?
The technique itself is not; its uses are. A supervised model that assesses the creditworthiness of individuals, screens job applications or decides on access to certain services may be high risk (Annex III). It must then comply with Article 10 on data quality and governance, obligations that apply from 2 December 2027 since the Digital Omnibus.
See also
Further reading
Supervised learning, Google for Developers (free introductory course)
Sources
- Apprentissage supervisé (supervised learning), definition, CNIL (French data protection authority). https://www.cnil.fr/fr/definition/apprentissage-supervise
- Supervised learning, Introduction to Machine Learning course, Google for Developers. https://developers.google.com/machine-learning/intro-to-ml/supervised
- Regulation (EU) 2024/1689 on artificial intelligence (AI Act), Article 10 and Annex III, EUR-Lex. https://eur-lex.europa.eu/eli/reg/2024/1689/oj
- Ouyang et al., Training language models to follow instructions with human feedback (InstructGPT), OpenAI, arXiv, 4 March 2022. https://arxiv.org/abs/2203.02155