Risk Modelling and Survival Analysis · Elementary principles of machine learning

Introduction to Machine Learning and Its Types

Updated 11 October 2026 · Fact-checked

Machine learning is a set of methods where a computer learns patterns from data to predict or group outcomes, with less reliance on a pre-specified model. The three main types are supervised (labelled outcomes), unsupervised (no labels) and reinforcement learning (learning from rewards). To answer questions, identify the data and the goal first.

Understand Introduction to Machine Learning and Its Types

Machine learning (ML) is a branch of data analysis in which algorithms learn patterns from data and use them to make predictions or find structure. The emphasis is on predictive accuracy on new data, not on explaining the process that created the data.

The contrast with traditional statistical modelling is one of emphasis. In a traditional approach, you choose a model form first, such as a linear regression or a Poisson distribution. You state assumptions, estimate parameters, and test hypotheses. Interpretation and inference matter. In ML, you often let a flexible algorithm find the form from the data. You judge it mainly by how well it predicts data it has not seen. The two overlap heavily: a linear regression is both a statistical model and a supervised ML method.

The main categories depend on the data you have.

Supervised learning: each record has inputs (features) and a known outcome (the label or target). The aim is to predict the outcome for new records. If the target is numeric, it is regression. If it is a category, it is classification. Examples: predicting claim size, or predicting whether a policyholder will lapse.
Unsupervised learning: there is no target. The aim is to find structure, such as groups (clustering) or fewer summary variables (dimension reduction). Example: grouping policyholders by behaviour.
Reinforcement learning: an agent takes actions in an environment and gets rewards or penalties. It learns a strategy that maximises long-term reward. Example: learning a trading or pricing policy by trial and error.

In practice, data are split into training data, used to fit the model, and test data, used to check it. A model that fits training data very well but predicts badly on new data is overfitted. This idea runs through the whole chapter and is the main reason ML needs care.

Key rules to remember

Supervised learning set-up: y = f(x) + ε; x are the features, y the target, f the unknown relationship to be learned, ε random error. Supervised learning estimates f.
Choosing the type of learning: Labelled target? Yes → supervised. No → unsupervised. Reward-driven actions? → reinforcement.; A decision rule, not a formula. Use it to classify any scenario.
Supervised sub-types: Numeric target → regression. Categorical target → classification.; A binary outcome such as lapse or no lapse is classification, even if the model outputs a probability.

How to solve Introduction to Machine Learning and Its Types questions

Use this method for any question that asks you to define, classify or compare machine learning approaches.

1Read the scenario and write down the data available: which variables are inputs and whether any outcome is recorded.
2Decide whether a labelled target exists. If yes, the problem is supervised. If no, it is unsupervised.
3If supervised, look at the target. Numeric means regression. Categorical means classification.
4If the scenario involves an agent taking actions and receiving rewards over time, name it reinforcement learning.
5State the goal in one line: predict, group, reduce dimensions or optimise a policy.
6Name a suitable method or example only if asked, and link it to the goal.
7Add a short comment on a limitation, such as overfitting, interpretability or data quality, if marks allow.

Quickest way: Label test

When to use it: Multiple-choice questions and short classification questions where time is tight.

Ask: is there a known outcome for each record?
If yes, supervised. Check whether the outcome is a number (regression) or a category (classification).
If no, unsupervised. Groups mean clustering. Fewer variables means dimension reduction.
If the words reward, action or agent appear, choose reinforcement learning.

Common mistakes in Introduction to Machine Learning and Its Types

Saying unsupervised learning has no data to learn from, or has no goal.
The word 'unsupervised' sounds like 'no information'.
Fix: Say it has no labelled target. It still has features and aims to find structure such as groups.
Calling a lapse or default prediction regression because it gives a probability.
The output looks numeric.
Fix: Look at the target. If it is a category such as lapse or not, it is classification.
Claiming ML and statistical modelling are completely separate.
Notes present them as opposites.
Fix: Say they overlap. The difference is mainly emphasis: prediction versus inference and assumptions.
Judging a model by its fit on training data alone.
A high training fit looks like success.
Fix: State that performance must be checked on data not used for fitting, to detect overfitting.
Confusing reinforcement learning with supervised learning.
Both involve feedback.
Fix: Supervised feedback is the correct label for each record. Reinforcement feedback is a reward for actions, often delayed.

Worked examples

Example 1

An insurer has 50,000 motor policies. For each it records age, vehicle type, region and the total claim amount in the last year (₹). It wants to predict claim amount for new policies. (a) Name the type of learning. (b) Name the sub-type. (c) Give one reason to test on separate data.

Show the solution

The claim amount is recorded for every policy, so a labelled target exists. This is supervised learning.
The target is a numeric amount, so the sub-type is regression.
The model is fitted on training data. A very flexible model may fit noise in that data.
Testing on separate data shows whether it predicts unseen policies well, which is the real aim.

Answer: (a) Supervised learning. (b) Regression. (c) Separate test data reveals overfitting and gives an honest measure of predictive performance.

Example 2

A life insurer has records of customers with age, premium, policy term and number of service calls. There is no outcome variable. It wants to find natural customer segments. (a) Name the type of learning and method. (b) Explain how this differs from predicting whether each customer lapses.

Show the solution

There is no outcome variable, so the problem is unsupervised.
Finding natural segments is grouping, so the method is clustering.
Predicting lapse needs a recorded lapse outcome for each customer, which makes it supervised classification.
Clustering finds structure without being told the answer. Classification learns from known answers to predict them for new customers.

Answer: (a) Unsupervised learning, using clustering. (b) Lapse prediction is supervised classification because it uses a labelled outcome. Clustering has no labels and only looks for groups in the features.

Exam tips

Start every classification answer by stating whether a labelled target exists. Examiners look for this first.
Give a short insurance example for each type. Claim size, lapse, segmentation and pricing policy are safe choices.
When asked to compare ML with traditional modelling, give one point on each side: interpretability and assumptions versus flexibility and prediction.
Always mention overfitting and out-of-sample testing when discussing performance. It is a standard mark.
In multiple-choice questions, read the target variable carefully before choosing regression or classification.

Practice questions from Elementary principles of machine learning

Introduction to Machine Learning and Its Types in other exams

The same ground in other exams, if you are preparing for more than one or want another angle on it.

Introduction to Machine Learning and Its Types: frequently asked questions

What is machine learning in actuarial science?

It is the use of algorithms that learn from data to predict outcomes or find structure, such as claim sizes, lapses or customer groups. Actuaries use it alongside traditional models. Judgement, interpretability and governance still matter.

What is the difference between supervised and unsupervised learning?

Supervised learning uses data with a known outcome for each record and learns to predict it. Unsupervised learning has no outcome and looks for patterns such as clusters. The presence of a labelled target is the deciding test.

Is regression a machine learning method?

Yes. Linear regression is a supervised learning method as well as a classical statistical model. Many ML methods extend it with more flexibility.

What is reinforcement learning?

It is learning by trial and error. An agent takes actions, receives rewards or penalties, and learns a policy that maximises long-term reward. It is covered at an introductory level in CS2.