The one question that decides which type you need
Ask whether your data contains the outcome you want to predict. If every past record has a known label, such as spam or not spam, or the final sale price of a house, you can teach a model to predict it. That is supervised learning.
If there is no outcome column and you want to understand the data itself, such as which customers behave alike or which transactions look unusual, you use unsupervised learning. Many real projects use both: clustering to explore, then a supervised model to predict.
Side by side comparison
How the two approaches differ in practice:
Data needed
Supervised needs labelled examples. Unsupervised works on unlabelled data, which is usually easier to get.
Typical goal
Supervised predicts a value or a class. Unsupervised finds groups, patterns or outliers.
Measuring success
Supervised compares predictions with known answers using accuracy, precision, recall or error. Unsupervised is judged by cluster quality scores and whether the result makes business sense.
Everyday examples
Supervised: spam filters, price prediction, churn risk. Unsupervised: customer segments, anomaly detection, topic grouping.
Algorithms you will meet first
Learn these before anything more advanced:
- Supervised regression, for predicting numbers: linear regression, decision trees, random forests and gradient boosting.
- Supervised classification, for predicting categories: logistic regression, k-nearest neighbours, decision trees and gradient boosting.
- Unsupervised clustering: k-means for simple round groups, hierarchical clustering, and DBSCAN for oddly shaped groups and noise.
- Unsupervised dimensionality reduction: PCA to compress many columns into a few, for visualisation or as a step before modelling.
- Unsupervised anomaly detection: isolation forest to flag records that look unlike the rest.
Choosing the right approach for a project
Work through these questions before you open a notebook.
Write the business question
For example, "which customers are likely to leave next month?" or "what kinds of customers do we have?"
Check for labels
If past data records who actually left, you can train a supervised model. If not, plan to collect labels or start with exploration.
Start simple
Fit a basic model first, such as logistic regression or k-means, to get a baseline you can explain.
Evaluate honestly
For supervised models, test on data the model never saw. For clusters, check whether each group is stable and means something to the business.
Combine when useful
Cluster labels can become a feature in a supervised model, and anomaly scores can be reviewed by people to create new labels.
Quick answers about supervised and unsupervised learning
Short answers to what learners search for most.
Is k-means supervised or unsupervised?
Unsupervised. K-means groups records by similarity without any labels, and you choose the number of groups.
Is linear regression supervised learning?
Yes. It learns from examples with known numeric outcomes, such as past sale prices, and predicts that number for new records.
What is semi-supervised learning?
It uses a small amount of labelled data together with a lot of unlabelled data. It is useful when labelling is slow or expensive, such as medical images.
Where does reinforcement learning fit?
It is a third type, where an agent learns by trial and error from rewards, as in game playing or robotics. It is less common in everyday business analytics.
Which should a beginner learn first?
Most courses start with supervised learning, because results are easy to measure against known answers, then add clustering and other unsupervised methods.
Continue with machine learning fundamentals
See where these methods sit in the Data Science syllabus, or read the related guides.
Let the labels decide
If you know the answer for past cases, train a supervised model and test it on unseen data. If you do not, explore with clustering first. Either way, start with the simplest model you can explain.