CertBrain
Courses
Home
Blog
Sign in
← Back to decks
Sign in to save your progress, vote, and build your own decks.Sign in

Supervised, Unsupervised and Reinforcement Learning

Machine learning comes in three main types, and the data you have usually decides which one fits. This lesson explains supervised, unsupervised and reinforcement learning with a typical use for each, then covers three terms the exam pairs with them: fit, bias and fairness.

9 cards
Study this deck

Supervised, Unsupervised and Reinforcement Learning

Machine learning comes in three main types, and the data you have usually decides which one fits. This lesson explains supervised, unsupervised and reinforcement learning with a typical use for each, then covers three terms the exam pairs with them: fit, bias and fairness.

Supervised learning

Supervised learning trains on labeled data: examples where the right answer is already known. The model learns to map inputs to those answers, then predicts the answer for new inputs. Two common jobs are classification, which predicts a category ("fraud" or "not fraud"), and regression, which predicts a number (next month's sales, a house price). If a scenario has historical examples with a known outcome column, supervised learning is the natural choice.

Unsupervised learning

Unsupervised learning works on unlabeled data and finds structure nobody defined in advance. The most common job is clustering: grouping similar items, such as customers with similar buying habits. Other uses include spotting unusual records (anomaly detection) and finding items that often occur together, such as products bought in the same basket. Choose it when you want to discover groups or patterns rather than predict a known answer. Because there's no answer key, a person still has to judge whether the groups it finds are useful.

A middle ground, semi-supervised learning, uses a small labeled set alongside a larger unlabeled one. It's useful when labels are expensive.

Reinforcement learning

Reinforcement learning (RL) learns by trial and error. A software agent takes actions in an environment and receives a reward or penalty for each outcome. Over many attempts, it learns which actions earn the most reward over time. No dataset of correct answers is needed, only a way to score what happened. RL suits sequences of decisions: robot movement, game playing, or tuning prices or ad placements as conditions change. It's also the idea behind training language models with human feedback, where people's ratings act as the reward.

Fit, bias and fairness

Fit describes how well a model captures the real pattern. An overfit model has memorised its training data, noise included, so it scores very well in training but poorly on new data. An underfit model is too simple to capture the pattern, so it does poorly on both.

Bias has two meanings. In the statistical sense, it is systematic error from a model that is too simple, which goes with underfitting. In the responsible AI sense, it means outcomes that are skewed against certain groups, often because the training data under-represents them. Fairness is the goal of a model treating people and groups equitably. It's checked by comparing a model's results across groups.

On the exam

Clue in the scenarioLearning type
Historical data with a known outcome to predictSupervised
No labels; find groups or unusual itemsUnsupervised
An agent, actions and a reward or scoreReinforcement

For fit, remember the pattern: great on training data and poor on new data means overfitting; poor on both means underfitting.

Quiz me on this
A streaming service wants to group its viewers into segments based on what and when they watch, so it can tailor its home page. It has no predefined segment names or labels. Which type of machine learning fits this goal?
  1. Supervised classification, because viewing history is already recorded
  2. Supervised regression, because watch time is a number
  3. ✓ Unsupervised learning, using clustering to find groups of similar viewers
  4. Reinforcement learning, rewarding the model whenever viewers keep watching longer
A warehouse wants its robots to learn the fastest routes between shelves as aisles get blocked and reopened. No dataset of correct routes exists, but every run can be scored automatically on travel time and near-collisions. Which learning approach fits BEST?
  1. ✓ Reinforcement learning, where each run's score acts as the reward
  2. Supervised learning on a small set of routes that staff label by hand
  3. Unsupervised clustering of past routes into similar groups
  4. Semi-supervised learning that combines labeled and unlabeled routes
A model scores 99% accuracy on the data it was trained on but only 70% on new data it has never seen. Which problem does this show?
  1. Underfitting, because the model is too simple to learn from the data
  2. ✓ Overfitting, because the model memorised its training data
  3. Demographic bias, because the model favours one group
  4. A well-fitted model that is ready for production
Supervised learning
Learning from labeled examples to predict a known kind of answer: a category (classification) or a number (regression).
Unsupervised learning
Finding structure in unlabeled data, such as clusters of similar items, unusual records or items that occur together.
Reinforcement learning
An agent learns by trial and error, taking actions and getting rewards or penalties, to maximise reward over time.
Overfitting
The model memorises its training data, so it scores well in training but poorly on new data.
Underfitting
The model is too simple to capture the real pattern, so it performs poorly on both training and new data.
Fairness (in AI)
The goal that a model's outcomes treat people and groups equitably. It's checked by comparing results across groups (subgroup analysis).