AI Glossary / AI and machine learning / Beginner
Supervised learning
Definition
Supervised learning trains a model using examples that pair inputs with known target outputs. The model learns a relationship from those examples so it can estimate an output for a new input whose answer is not yet available.
What it means
An input might describe a delivery, and its target might record whether it arrived late. Another task might use property characteristics as inputs and a recorded sale price as the target. The first predicts a category; the second predicts a number.
The word supervised refers to the availability of target answers during learning. It does not mean that a person watches every calculation or approves every prediction. Targets may come from operational records, expert annotations or other data collection processes.
The goal is useful performance on new cases, not perfect recall of old answers. A model can memorize its training examples and still fail when circumstances change.
For a business, the central questions are whether the target represents the real objective and whether the inputs would be available at the time a decision is made. Those choices often matter more than selecting a sophisticated algorithm.
How it works
The team defines an example, its input features and the target to predict. It collects records and separates data for training, development and final evaluation in a way that reflects actual use.
During training, the model produces estimates. A loss function measures disagreement with the known targets, and a learning procedure adjusts the model to reduce that disagreement. Different algorithms perform this adjustment in different ways.
Development data helps select settings and compare approaches. A separate test set estimates performance after those decisions are made. If the task involves future events, time-based separation may be needed to avoid learning from the future.
After deployment, the team monitors both inputs and outcomes. Historical relationships may weaken when customers, products or operations change. Repeated evaluation is necessary even when the initial test result was strong.
Three practical examples
- Email routing: messages paired with reviewed department labels teach a classifier to suggest where new enquiries belong.
- Demand estimation: past product and calendar information paired with observed sales supports numerical predictions for planning.
- Quality inspection: photographs paired with verified defect labels train a system to flag products for further inspection.
Think of it this way
Think of practicing with worked exercises. You attempt an answer, compare it with the supplied answer and use the difference to improve your method.
The analogy has a limit. A person can ask whether an answer key is wrong or whether a question is misleading. A learning system may absorb a bad label without understanding its origin. Reviewing the examples and defining the task remain human responsibilities.
Worked example
Building a delivery-delay predictor
Suppose an invented project has 10,000 completed deliveries. Each example includes information known when dispatch was planned, such as route, parcel size and scheduled day. Its target records whether the delivery later exceeded the agreed deadline.
The team uses 7,000 examples for training, 1,500 for development and 1,500 for testing, with appropriate chronological separation. These numbers illustrate roles, not a mandatory split ratio.
If 150 test deliveries were late and the model flags 120 of them, recall for late deliveries is 120 / 150 = 80%. The team still needs to count on-time deliveries incorrectly flagged before judging usefulness.
Including the actual arrival time among the inputs would make the historical problem artificially easy. That value is unavailable when dispatch is planned, so it would leak the answer rather than support a valid early prediction.
Illustrative case study
A fictional maintenance company trains a model to predict repeat visits using completed job records. Early results look excellent because a field called follow-up booked often appears after the technician has already recognized the need for another visit.
When the model is used before the first visit, that information is missing. The team rebuilds the dataset around a clear prediction time and includes only information available then. It compares the revised model with a simple historical baseline.
The lesson is that a good-looking training task can differ from the operational decision. Supervision is useful only when the examples represent the question the deployed system will actually face.
Common misconceptions and mistakes
- Supervised means a human approves every output. It describes training with target answers, not an approval process.
- More labeled records always improve the model. Incorrect, duplicated or unrepresentative labels can make learning worse.
- A prediction explains what causes the outcome. Predictive associations do not by themselves establish causation or identify an effective intervention.
Business use
Supervised learning suits repeated decisions where past examples and meaningful outcomes are available. It can support prioritization, forecasting, classification and estimation.
Start with a specific decision and a measurable baseline. Include the costs of mistakes, annotation, monitoring and human review. Ask whether the target can be observed reliably and quickly enough to assess performance after deployment. A delayed outcome can make ongoing evaluation harder.
People also ask
Does it require manually labeled data?
No. Targets may already exist in business records, although their meaning and reliability still need checking.
Is regression supervised learning?
Yes. Regression predicts numerical targets; classification predicts categories.
Can supervised learning work with text or images?
Yes. The inputs can take many forms as long as the chosen method and preparation process can represent them appropriately.
Check your understanding
A model predicts next month's cancellations using a field populated only after a cancellation happens. Why is the test result suspect?
Answer: The field reveals information unavailable at the prediction time. The evaluation contains leakage and may greatly overstate real performance.
Related terms
Learn next
Read training set and test set. For one proposed model, record exactly when the prediction is made and which input fields exist at that moment.
More in AI and machine learning
- Sufficient dimension reduction - Sufficient dimension reduction seeks a smaller representation of predictors that preserves all their information about a specified response.
- Superintelligence ban - A superintelligence ban is a proposed prohibition on developing, deploying, or otherwise enabling AI systems that exceed a defined superintelligence threshold.
- Supervised Clustering - Supervised clustering groups objects using guidance from known labels.
- Support vector machine - A support vector machine, or SVM, is a learning method that can build a decision boundary between classes while balancing separation against errors.
- SUPS - SUPS stands for synaptic updates per second, a throughput measure used in neural simulation and some neural computing contexts.
- Surrogate model - A surrogate model is a simpler or cheaper model used to approximate the behavior of another system, model, experiment, or computational process.
Sources
Google, Introduction to Machine Learning, Supervised Learning: https://developers.google.com/machine-learning/intro-to-ml/supervised. Related glossary definition consulted at https://developers.google.com/machine-learning/glossary#supervised_machine_learning. Business scenarios and arithmetic are original.
Previous: Supervised fine-tuning Next: Supervised principal component analysis