Types of Machine Learning

1.2 — Types of Machine Learning

Recall first. You receive 10,000 customer records. In one project every record has a known “left / stayed” outcome; in another, no outcome is provided. Which project can directly learn from labels, and what could the second project discover?

The classification question

The most useful first question is:

What feedback does the learner receive?

• Supervised learning: examples include a target or label.
• Unsupervised learning: examples have no supplied target; the algorithm looks for structure in the inputs.
• Reinforcement learning: an agent acts in an environment and receives rewards or penalties over time.

Semi-supervised and self-supervised learning are important variants, but they do not replace the basic distinction: identify what information is available at learning time and what objective is being optimized.¹²

Supervised learning

A supervised dataset contains pairs (xᵢ, yᵢ). The learner uses the known yᵢ values to learn a function that predicts y for new x.

Classification

Classification predicts a discrete class:

• spam / not spam;
• defective / acceptable;
• disease present / absent;
• one of several recognized species.

The output may be a class label or a score/probability that is later converted into a class using a threshold.

Regression

Regression predicts a numeric quantity:

• house price;
• temperature;
• travel time;
• exam score.

A numeric output does not guarantee that regression is appropriate. The target must have a meaningful numeric interpretation and the loss must match the decision problem.

scikit-learn places linear models, trees, support-vector methods, and neural-network methods into supervised-learning families because they learn from targets supplied with the training examples.³

Unsupervised learning

An unsupervised dataset contains xᵢ but no supplied yᵢ. The goal is not “predict the missing label” unless a later task supplies that interpretation. Common objectives include:

• Clustering: group similar examples.
• Dimensionality reduction: represent data with fewer variables while preserving selected structure.
• Density or anomaly analysis: identify regions of common or unusual behavior.
• Representation learning: construct useful features from raw inputs.

The lack of labels does not mean the method has no objective. It still optimizes a criterion such as within-cluster distance or reconstruction error. The challenge is that a mathematically neat structure may not be useful to a person or application.

Reinforcement learning — the boundary

In reinforcement learning, an agent chooses actions, observes consequences, and receives rewards. The feedback can be delayed: an action may be good because of what it enables many steps later. This differs from supervised learning, where each training example normally comes with a target answer.

Example: a robot learns a navigation policy from rewards for reaching a destination safely. It is not simply given a correct action for every state.

Reinforcement learning is adjacent to the Module I foundations; the core exam distinction is the feedback signal, not memorizing a list of algorithms.

Related variants

• Semi-supervised learning: uses a smaller labelled set together with a larger unlabelled set.
• Self-supervised learning: creates a learning signal from the data itself, such as predicting a hidden part of an input.
• Online learning: updates a model as examples arrive rather than fitting only once to a fixed batch.

These labels describe data or training arrangements. A particular system can combine them with classification, regression, or representation learning.

A decision table for choosing the type

The table is a starting point, not a guarantee. A project can fail because its labels are unreliable, the objective is wrong, or the data available at deployment differs from the training data.

Worked example — classify the problem

A factory records vibration measurements from motors. It has:

• Dataset A: vibration readings plus a label “failed within 30 days / did not fail.”
• Dataset B: vibration readings only, with no failure labels.
• Dataset C: readings from a robot that receives a reward when it reaches a maintenance station efficiently.

Answer: Dataset A supports supervised classification. Dataset B supports unsupervised exploration, such as grouping operating patterns, but clusters are not automatically failure labels. Dataset C is a reinforcement-learning setting because the feedback comes from actions and rewards.

Exercise

Name the learning type and explain your choice:

1. Grouping customers into segments with no predefined segment names.
2. Predicting the monthly sales amount from past labelled records.
3. Learning to play a game using wins, losses, and intermediate rewards.
4. Predicting whether a tumour is benign from labelled clinical examples.

Exam lens

Three-way comparison: write the available feedback, objective, and example for supervised, unsupervised, and reinforcement learning.

Common traps:

• Unsupervised does not mean “random” or “without an objective.”
• Classification predicts categories; regression predicts numbers.
• A cluster is not automatically a real-world class.
• A probability score is not necessarily the final class label; a decision threshold may be applied later.

Rapid revision

• Can I decide the learning type from the feedback signal?
• Can I distinguish classification from regression?
• Can I explain why an unlabelled cluster is an interpretation, not a guaranteed truth?
• Can I distinguish supervised labels from reinforcement rewards?

Sources