The main types of machine learning are supervised, unsupervised, semi-supervised, self-supervised, and reinforcement learning. Each one teaches a model in a different way. The biggest difference is the kind of feedback the model gets while it learns.
Some models learn from examples with known answers. Others find patterns on their own, use a mix of labeled and unlabeled data, build practice tasks from raw data, or learn from rewards and penalties. The right choice depends on your goal and the data you have.
There is no single list of machine learning types used by every source. IBM’s overview of machine learning types lists self-supervised learning as a separate approach, while some references group it under semi-supervised learning. This article explains it separately because it uses a distinct way to create training targets from raw data.
The 5 Types of Machine Learning at a Glance
These types are not five separate boxes. They overlap, and one project can use more than one. A simple way to tell them apart is to ask where the model gets its feedback.
| Type | What the model learns from | Main goal |
|---|---|---|
| Supervised learning | Labeled examples | Learn to predict answers for new data |
| Unsupervised learning | Unlabeled data | Find patterns or groups |
| Semi-supervised learning | A mix of labeled and unlabeled data | Learn when only some examples have labels |
| Self-supervised learning | Targets made from the data itself | Learn from raw data without hand-made labels |
| Reinforcement learning | Rewards and penalties from actions | Learn which actions lead to better results |
Supervised Learning
Supervised learning trains a model on examples that already have the correct answers. Those answers are called labels. Say a real estate company wants to estimate home prices. It has records of past sales, and each record lists the square footage, bedrooms, bathrooms, ZIP code, and final sale price.
The home details are the input, and the sale price is the label. The model studies many records and learns how the features relate to price. After training, it can estimate the price of a home it has never seen.
Regression and Classification
Supervised learning handles two main kinds of tasks. Regression predicts a number, such as a home’s sale price, next month’s electricity use, or a delivery time. Classification predicts a category, such as spam or not spam, cat or dog, or fraud or not fraud.
A classification model often gives a probability. For example, it might say an email has an 85% chance of being spam. The system then uses a cutoff, called a threshold, to make the final call.
When Supervised Learning Is a Good Fit
Use it when you have examples with known answers and want similar predictions on new data. It is common in fraud detection, sales forecasting, medical image analysis, and spam filtering. Label quality matters a lot. If past records contain wrong prices or wrong categories, the model can learn those mistakes.
Unsupervised Learning
Unsupervised learning works with data that has no ready-made answers. Instead of being told what each example means, the model looks for patterns, similarities, or groups. Picture an online store with thousands of customers. It knows what people buy, how often they shop, and how much they spend, but it has no labels like “budget shopper” or “frequent buyer.”
An unsupervised method can sort those customers into groups with similar habits. The store can then study each group and decide how to serve it.
Clustering and Dimensionality Reduction
Clustering groups similar data points. A store might find shoppers who buy low-cost items often, shoppers who mostly buy during holiday sales, and shoppers who make big purchases less often. People then review the groups and decide what they mean.
Dimensionality reduction makes large datasets easier to handle. A dataset can have hundreds of features, and some features may overlap or carry less useful information than others. Methods such as principal component analysis (PCA) keep most of the useful variation with fewer features.
When Unsupervised Learning Is a Good Fit
Use it when you want to explore data and do not have known answers for every record. Common uses include customer grouping, spotting unusual activity, and exploring large datasets. Keep in mind that a group found by a model is not always meaningful. A cluster may look clear in the data but still be useless for the business, so a person needs to check the results.
Semi-Supervised Learning
Semi-supervised learning uses both labeled and unlabeled data. It helps when labels are slow, costly, or need expert knowledge, but raw data is easy to collect. A hospital is a good example. It may have thousands of medical images, but only a small share have been reviewed and labeled by specialists.
A semi-supervised method trains on the small labeled set and also draws on the unlabeled images. The labels give known answers, and the unlabeled images add more information about what the data looks like.
How It Works
The exact steps depend on the method. One common approach looks like this:
- Train a first model on the small labeled set.
- Use that model to predict labels for some unlabeled examples.
- Keep the predictions that meet a chosen confidence level.
- Add those predictions to the training data and train again.
The predicted labels are called pseudo-labels. They are not guaranteed to be right. If the model treats a wrong guess as a true answer, the error can spread in later rounds.
When Semi-Supervised Learning Is a Good Fit
It works well in image classification, speech tasks, and other areas where unlabeled data is plentiful but good labels are expensive. The unlabeled data still has to fit the task. If a model is learning to tell cats from dogs, a pile of car photos will not help.
Self-Supervised Learning
Self-supervised learning lets a model make its own learning task from raw data. No one has to label each example by hand. This matters most for language and image models, where huge amounts of raw data exist but hand labeling would take far too long. Google’s machine learning glossary explains how self-supervised learning creates training targets from unlabeled data.
Here is a simple text example. The model sees the sentence “The dog chased the ___ across the yard” and has to guess the missing word. The original text holds the answer, so no person had to write a label. After many rounds on many sentences, the model picks up patterns in how words are used and how they relate. Image models can do something similar by rebuilding parts of a picture that were hidden during training.
Self-Supervised Learning in Language Models
Many large language models use self-supervised tasks during pretraining. A common one is predicting the next token in a sequence. A token can be a word, a piece of a word, or another unit of text.
Pretraining is often not the last step. A model may later go through supervised fine-tuning or other training to make it more useful for specific jobs.
How It Differs From Unsupervised Learning
Both approaches can start with unlabeled data, which is why people mix them up. The difference is the goal. Self-supervised learning builds a target from the data, such as a hidden word, and trains the model to predict it. Unsupervised learning looks for groups or structure without that kind of target.
Reinforcement Learning
Reinforcement learning trains a system to make decisions by taking actions and getting feedback. The feedback comes as rewards or penalties. The system is not handed the right action for every situation. It tries things, sees what happens, and slowly learns which choices lead to better results.
A reinforcement learning setup has four main parts:
- Agent: the learner that makes decisions.
- Environment: the world or system the agent works in.
- Action: a choice the agent makes.
- Reward: feedback that shows how good the outcome was.
Take a game-playing program. The agent picks a move, the game responds, and the agent gets feedback. A move that helps it win earns a reward, and a move that leads to a loss earns a penalty. After many rounds, the agent can learn a strategy that earns more reward.
Where Reinforcement Learning Is Used
It shows up in robotics, game playing, resource management, and control systems. A robot can learn how to move or pick up objects. A control system can learn which actions help it reach a target.
It is not the best tool for every prediction problem. It fits best when a system must choose actions over time, and each action changes what happens next. That is also what sets it apart from supervised learning. A spam filter learns from emails already marked as spam or not spam, while a game agent learns from the results of its own moves.
How the Types of Machine Learning Compare
The table below puts the five approaches side by side.
| Approach | Training data | How it learns | Example |
|---|---|---|---|
| Supervised | Labeled | Compares its predictions with known answers | Predicting home prices |
| Unsupervised | Unlabeled | Finds patterns and similarities | Grouping customers |
| Semi-supervised | Labeled and unlabeled | Uses both kinds of examples | Classifying medical images |
| Self-supervised | Unlabeled, with targets made from the data | Predicts a hidden part of the input | Predicting the next word |
| Reinforcement | Feedback from actions | Learns from rewards and penalties | Learning to play a game |
Semi-supervised and self-supervised learning sound alike, but their training signals are different. Semi-supervised learning relies on a small set of human labels plus extra unlabeled data. Self-supervised learning makes its own targets and does not need people to label each example. The two can also work together. A project may start with self-supervised pretraining and then use labeled data to fine-tune the model for one task.
How to Choose the Right Type of Machine Learning
No single type works best for every problem. Start with the question you want to answer, then look at your data.
| Your situation | Approach to consider | Example |
|---|---|---|
| You have examples with known answers | Supervised learning | Predicting sale prices |
| You want to find groups or hidden patterns | Unsupervised learning | Grouping shoppers by behavior |
| You have a small labeled set and a larger related unlabeled set | Semi-supervised learning | Classifying a large image collection |
| You have lots of raw data and can build a learning task from it | Self-supervised learning | Pretraining a language model |
| A system must make choices and learn from the results | Reinforcement learning | Training a game-playing agent |
These four questions can help you narrow it down:
- Do you have correct answers for your examples? If yes, supervised learning is often a good place to start. If not, decide whether you want to find patterns or build a learning task from the data.
- Is labeling hard or costly? If you have a few labels and many related unlabeled examples, try semi-supervised learning. If you have a large pile of raw data, look at self-supervised learning.
- Does the system need to make decisions over time? If so, reinforcement learning may fit better than a standard prediction method.
- How will you know it works? A model can find patterns that look interesting but do not solve your problem. Decide how you will measure success before you pick an approach.
Common Mix-Ups About Machine Learning Types
A few related terms often get confused with the five types above.
Is Deep Learning a Type of Machine Learning?
Not in the same sense. Deep learning is a way to build models using neural networks with many layers. It is not a learning approach on the same level as supervised or reinforcement learning. A deep learning model can be trained with supervised, self-supervised, or reinforcement methods, among others.
Is Generative AI a Type of Machine Learning?
No. Generative AI describes systems that create new content, such as text, images, audio, or code. These systems are built with machine learning models, and their training can include self-supervised learning, supervised fine-tuning, and reinforcement learning.
Can One Model Use More Than One Approach?
Yes. A language model might first learn from a huge amount of text through self-supervised training. It might then be fine-tuned on examples with human-written answers, and later trained further using feedback to make its replies more helpful. This is why the five types are better seen as different ways to learn, not as separate groups.
Are Learning Types and Algorithms the Same?
No. A learning type describes how a model gets its training information. An algorithm is the specific method used to do the learning. Supervised learning is a type, and linear regression is one algorithm you can use for a supervised task.
Frequently Asked Questions About the Types of Machine Learning
What are the five main types of machine learning?
They are supervised, unsupervised, semi-supervised, self-supervised, and reinforcement learning. They differ in how a model gets its training information and feedback.
Which type of machine learning is used most?
No single type leads in every field. Supervised learning is widely used for tasks like price prediction, spam detection, and fraud detection. Self-supervised learning plays a big role in language and image models.
Which type is best for beginners?
Supervised learning is a common starting point because the examples are easy to follow. You can use labeled house records to predict prices, or labeled emails to spot spam. Unsupervised learning is a good second step for seeing how models find patterns on their own.
Is reinforcement learning used in real life?
Yes. It is used in robotics, game playing, and control systems, where an agent needs to choose actions and learn from what happens next.
Is deep learning supervised or unsupervised?
It can be either, depending on how the model is trained. Deep learning describes the model’s structure, while supervised and unsupervised describe how it learns.
Final Thoughts
The five types of machine learning differ mainly in the feedback a model learns from. Labeled examples point to supervised learning, hidden patterns point to unsupervised learning, and a mix of labeled and unlabeled data points to semi-supervised learning. Raw data with a built-in task suits self-supervised learning, and decisions that earn rewards suit reinforcement learning.
In real projects, the lines often blur. Look at your goal and your data first, and the right approach usually becomes clear.
You Might Also Like:
How Does Machine Learning Work? A Step-by-Step Guide
I’m Emma Rose, the founder of Punstation. I cover technology, digital trends, business, and online tools, creating practical guides and thoughtful reviews to help readers understand their options and make informed decisions. For more on how I approach my content, see our Editorial Standards.