Two ways of learning
Machine learning methods are often grouped by the kind of guidance they receive during training. The two best-known groups are supervised and unsupervised learning. The difference comes down to one question: do the training examples come with the right answers attached?
Supervised learning: learning with an answer key
In supervised learning, each training example includes a label, the outcome the system should learn to predict. A teacher, in effect, is supplying the correct answer for every example.
Typical examples:
- Emails marked as spam or not spam, used to build a spam filter.
- Photos tagged with what they show, used to train an image recognizer.
- Past house sales with their final prices, used to estimate the price of another house.
- Loan records marked as repaid or not repaid, used to flag likely risk.
The system compares its guesses with the labels and corrects itself until it predicts well. Afterward, it can be applied to new, unlabeled cases. The process is described in more general terms in the walkthrough of how machine learning learns from data.
Supervised tasks come in two main flavors.
- Classification predicts a category, such as spam or not spam.
- Regression predicts a number, such as a price or a temperature.
Strengths
- The goal is clear: predict the label.
- Success is straightforward to measure, by checking predictions against known answers held back from training.
- It is widely used and well understood in many business settings.
Limits
- Labels cost effort. Someone must mark each example, which can be slow, expensive or require expertise. Poor or inconsistent labels harm results.
- Labels can embed human error or bias. If past decisions were skewed, a system trained to copy them may repeat the problem, as in concerns about algorithmic bias.
- It only knows what it was taught. It can predict only categories and relationships represented in its labeled data.
Unsupervised learning: finding structure unaided
In unsupervised learning, the examples have no labels. The system is given raw data and asked to find structure by itself. There is no answer key, so the output is less about “right” and “wrong” and more about whether the discovered structure is useful.
Common tasks:
- Clustering groups similar items together. A retailer might group customers by purchasing behavior without defining the groups in advance.
- Dimensionality reduction compresses many measurements into a few that capture the main variation, which helps with visualization and with simplifying data.
- Anomaly detection identifies items that look unlike the rest, such as unusual network activity or odd transactions.
- Association discovery finds things that tend to occur together, such as items often bought in the same basket.
Strengths
- No labeling needed, so it can use large amounts of raw data.
- It can reveal patterns people did not think to look for.
- It is useful for exploring unfamiliar data before deciding on a question to ask.
Limits
- Results are open to interpretation. A clustering method will always return groups, but whether those groups mean anything depends on human judgment.
- Hard to evaluate. Without known answers, quality is judged by usefulness, not by a simple accuracy figure.
- Sensitive to choices. The number of groups requested, or how similarity is measured, can change the outcome considerably.
- Patterns can be accidental. Large datasets contain coincidences that look like structure.
A quick comparison
| Feature | Supervised | Unsupervised |
|---|---|---|
| Training data | Examples with labels | Examples without labels |
| Typical goal | Predict a known outcome | Discover grouping or structure |
| Evaluation | Compare with known answers | Judge by usefulness and review |
| Main cost | Creating labels | Interpreting results |
| Example | Spam detection | Customer grouping |
Choosing between them
The choice usually follows from the problem. If there is a specific outcome to predict and labeled history exists, supervised learning is the natural fit. If the aim is to explore data, segment it or flag oddities, and no labels exist, unsupervised methods are suitable.
In practice the two are often combined. A team might cluster unlabeled data to understand it, then label a sample from each cluster to train a supervised model. Anomaly detection tools sometimes learn what normal looks like from unlabeled data and then raise flags for review.
Hybrids and related styles
Semi-supervised learning uses a small amount of labeled data together with a larger unlabeled set, which can reduce labeling effort.
Self-supervised learning creates its own labels from the data itself. Large language models are a well-known example: the “label” for each stretch of text is simply the piece of text that comes next. This blurs the line between the two categories and is one reason the topic has become central to modern AI.
Reinforcement learning is a separate style where a system learns from rewards for actions, not from labels or pure pattern discovery.
Why data quality decides outcomes
Whichever approach is chosen, the results depend on the data supplied. Mislabeled examples undermine supervised models, and unrepresentative samples distort unsupervised ones. The article on training data discusses how these choices shape behavior, and the idea of data quality applies equally to both approaches. Neither style is inherently better; each is a tool suited to different questions, and neither removes the need for testing and human review of what the system produces.