Supervised versus unsupervised learning, the confusion matrix, calculating accuracy, and how ML is applied in the real world.
Written and reviewed by the IK Learning team
This chapter introduces Artificial Intelligence and Machine Learning, explaining the difference between supervised and unsupervised learning, how to evaluate a model using a confusion matrix and accuracy, and real-world applications of AI and ML.
What this chapter is about, and why it matters
Class 9 introduced what AI and machine learning are. This chapter is about how machine learning actually works and — more importantly — how you tell whether it is working well.
The central division is between supervised learning, where the training data comes with correct answers attached, and unsupervised learning, where it does not. Nearly every question in this chapter depends on placing a scenario correctly on that line.
The confusion matrix is where students lose the most marks, almost always by mixing up false positives and false negatives. Learn to read the words literally — "false positive" means the system said positive and was false — and it becomes mechanical.
The skills this chapter is assessed on
8 core ideas — each with its definition and a separate worked example
Definition
Machine learning is a branch of artificial intelligence in which a system learns patterns from data and uses them to make predictions or decisions, without being explicitly programmed with the rules for the task.
Example
A system that predicts a student's final grade from their attendance and test scores is not given a formula. It is shown hundreds of past students' records and works out the relationship itself.
Definition
Supervised learning is a machine learning approach in which the model is trained on data where each example is already labelled with the correct answer, so the model learns to map inputs to known outputs.
Example
Training an email filter on 50,000 emails each already marked "spam" or "not spam". The label is the supervision — the model checks its guess against the known answer and adjusts.
Detailed Explanation
The word "supervised" refers to the labels, not to a person watching. If the training data contains the right answers, it is supervised, and the practical cost is that someone had to produce those labels first.
Definition
Unsupervised learning is a machine learning approach in which the model is trained on data with no labels, and must discover structure, groupings or patterns in the data by itself.
Example
Giving a shop's system a year of purchase records with no categories attached. It groups customers into clusters with similar buying habits. Nobody defined the groups in advance — the model found them.
Detailed Explanation
The output of unsupervised learning needs human interpretation. The system can report that three distinct groups exist; deciding that group two are "weekly bulk shoppers" is a judgement a person makes afterwards.
Definition
Training data is the set of examples used to teach a machine learning model. A label is the known correct output attached to a training example in supervised learning.
Example
For a model recognising handwritten digits, the training data is thousands of digit images, and each image's label is the number a human confirmed it shows.
Detailed Explanation
Data is normally split so that some examples are held back for testing. Evaluating a model on the same data it learned from is like marking a test using the answer sheet the student memorised — it measures recall, not understanding.
Definition
A confusion matrix is a table used to evaluate a classification model, showing the counts of true positives, true negatives, false positives and false negatives produced when its predictions are compared with the actual correct answers.
Example
A spam filter tested on 100 emails:
Actually Actually
Spam Not Spam
Predicted Spam 40 (TP) 5 (FP)
Predicted Not Spam 3 (FN) 52 (TN)Detailed Explanation
Read each term literally, in two words. "False positive" = predicted positive, and that was false → a genuine email sent to the spam folder. "False negative" = predicted negative, and that was false → spam that reached the inbox. Once you read them this way they stop being interchangeable.
Definition
Accuracy is the proportion of all predictions that were correct, calculated as the number of true positives plus true negatives, divided by the total number of predictions.
Example
From the matrix above: (40 + 52) ÷ 100 = 0.92, or 92% accuracy. Ninety-two of the hundred emails were classified correctly.
Detailed Explanation
Accuracy is misleading when one outcome is rare. If only 1 email in 100 were spam, a model that simply labelled everything "not spam" would be 99% accurate and completely useless. Always ask what the split of the data was before trusting an accuracy figure.
Definition
The two error types have different real-world consequences, so the acceptable balance between them depends on the application rather than on mathematics alone.
Example
For a medical screening test, a false negative tells a sick patient they are healthy — potentially fatal. A false positive causes worry and a second test. Here the system should be tuned to minimise false negatives, even at the cost of more false positives.
Detailed Explanation
For a spam filter the priority reverses: losing an important email in the spam folder (false positive) is worse than seeing one spam message in the inbox. This is why "which error matters more?" is an application question, and why exam answers should name the context.
Definition
Linear regression is a supervised learning technique that models the relationship between variables by fitting a straight line through the data, allowing a continuous numeric value to be predicted from an input.
Example
Plotting hours studied against marks scored for 200 students and fitting a line through the points. The line can then estimate the marks for a student who studied a number of hours not present in the original data.
Detailed Explanation
Regression predicts a number (marks, price, temperature); classification predicts a category (spam or not spam). Questions asking which technique fits a problem are really asking whether the answer is a quantity or a label.
How to lay the answer out so method marks are earned
Question: A model tested on 200 samples produced 70 true positives, 100 true negatives, 20 false positives and 10 false negatives. Calculate its accuracy and comment on the result.
Answer
Accuracy = 85%. Because the two classes are reasonably balanced, 85% is a meaningful figure here. Had 195 of the 200 samples been negative, the same accuracy would have been close to worthless.
The same ideas, outside the syllabus
Video and shopping platforms use both approaches: unsupervised clustering groups users with similar tastes, and supervised models predict whether a specific user will click a specific item. The recommendation you see is the output of both working together.
Screening tools are deliberately tuned towards more false positives to reduce false negatives, because a follow-up test is far cheaper than a missed diagnosis. This is the confusion matrix influencing a life-and-death design decision.
A school could use linear regression on attendance and past scores to identify students likely to struggle, so support arrives before the exam rather than after it. The same prediction also raises fairness questions worth discussing.
Errors that cost marks in this chapter, and the correction for each
Mistake
Swapping false positives and false negatives.
Correct Approach
Read the phrase literally: the second word is what the model predicted, the first says whether that prediction was right. False positive = predicted positive, was wrong.
Mistake
Defining supervised learning as "learning with a teacher watching".
Correct Approach
The supervision is the labelled data — correct answers supplied with the training examples. No person watches the training.
Mistake
Treating a high accuracy figure as proof that a model is good.
Correct Approach
On imbalanced data, accuracy is misleading. A model predicting the majority class every time can score 99% and be useless.
Mistake
Testing a model on the same data used to train it.
Correct Approach
Data must be split into training and testing sets, or the evaluation measures memorisation rather than genuine performance.
Mistake
Confusing regression with classification.
Correct Approach
Regression predicts a continuous number. Classification predicts a category. Ask what kind of answer the problem needs.
Mistake
Calculating accuracy using only the true positives.
Correct Approach
The formula is (TP + TN) ÷ total. Correct negatives are correct predictions too, and omitting them halves your answer.
Technique specific to this chapter
The whole chapter in one screen — read this the night before
Words used in this chapter, defined plainly
Now test yourself on the concepts above. Collapse the answers to make it a real practice run.
9 MCQs — pick an option to check yourself, then read why the answer is right
1How is Artificial Intelligence (AI) best described?
Correct answer: B — Machines performing tasks like humans
Artificial Intelligence enables machines to perform tasks that normally require human intelligence, such as reasoning and learning.
2Which of these counts as an example of AI?
Correct answer: A — A chatbot replying to your questions
A chatbot that understands and answers questions uses AI, unlike simple everyday devices that follow fixed actions.
3How is Machine Learning (ML) best described?
Correct answer: B — Teaching machines to learn from data like humans
Machine Learning is about teaching machines to learn patterns from data rather than being explicitly programmed for every task.
4Which app uses ML to recommend videos?
Correct answer: C — YouTube
YouTube uses machine learning to study viewing habits and recommend videos a user is likely to enjoy.
5In supervised learning, the data used is:
Correct answer: C — Labeled with correct answers
Supervised learning trains on data that is labeled with the correct answers so the model can learn to predict them.
6Which of these is an example of unsupervised learning?
Correct answer: D — Grouping students by interest without labels
Unsupervised learning finds hidden patterns in unlabeled data, such as grouping students by interest without any given labels.
7Google Assistant is an example of:
Correct answer: C — A smart assistant using AI
Google Assistant is an AI-powered smart assistant that understands voice commands and responds intelligently.
8AI is chiefly applied to:
Correct answer: D — Smart decision making and automation
AI is mainly applied to smart decision-making and the automation of tasks that would otherwise need human effort.
9A chatbot on JazzCash is an example of:
Correct answer: C — AI in business
A chatbot on JazzCash is an example of AI in business, using automation to assist and respond to customers.
8 short-answer questions
3 in-depth answers
4 high-priority questions
6 quick answers to common questions about this chapter
9 questions with the answers hidden — check what you actually remember
You have just read the explanations above. This checks whether they stuck. The answers stay hidden until you finish, so it is closer to exam conditions than scrolling through the notes again.
Your score is saved in this browser only. No account, nothing sent anywhere.