Machine Learning Techniques and Use Cases
Supervised learning: classification and regression
The story: A new fruit seller learns from an experienced one, who points at each fruit and says what it is. Soon the new seller can answer two kinds of questions alone. "Is this mango ripe or not ripe?" has a fixed set of answers. "How much will this watermelon weigh?" has an answer that can be any number.
In AI/AWS terms: Learning from someone who knows the answers is supervised learning: training on labeled data, where the labels act as the supervisor.
- "Ripe or not ripe" is classification: predicting a category. Use cases: fraud detection, image classification, customer retention (churn), diagnostics.
- "How many kilos" is regression: predicting a continuous number. Use cases: house prices, sales forecasts, demand estimates.
For the exam: Classification predicts a category. Regression predicts a number. Both are supervised.
Unsupervised learning: clustering and dimensionality reduction
The story: A new shop owner has a year of receipts but nobody to explain the customers. Looking at the receipts, they notice natural groups: early-morning coffee buyers, weekend family shoppers, late-night snack buyers. Separately, their shop report has 200 columns, so they boil it down to the five that really tell the story.
In AI/AWS terms: Finding structure with nobody giving answers is unsupervised learning: training on unlabeled data.
- Grouping customers is clustering: putting similar data points together. Use cases: customer segmentation, targeted marketing, recommendation systems.
- Boiling 200 columns down to five is dimensionality reduction: cutting the number of features while keeping the important information, for example before visualization or to speed up training.
For the exam: Clustering finds groups in unlabeled data. Dimensionality reduction cuts the number of features.
Reinforcement learning
The story: Learning to ride a bike. Nobody can hand you a list of exact muscle movements. You know what success looks like (staying up and moving forward), you try, you wobble, you fall, and slowly your body works out what keeps you upright.
In AI/AWS terms: In reinforcement learning, an agent learns by trial and error in an environment, guided by rewards and penalties. Use it when you know what a good outcome is but not the path to get there.
In AWS DeepRacer, the agent is the car, the environment is the track, the actions are throttle and steering, and the reward encourages finishing fast without leaving the track.
For the exam: Reinforcement learning = agent, environment, actions, rewards. Use it when you know the goal but not the steps.
Quick decision guide
The story: Before picking a tool from the toolbox, ask what you have and what you want out of it.
In AI/AWS terms:
| You want to... | Technique |
|---|---|
| Predict a label from labeled examples | Classification |
| Predict a number from labeled examples | Regression |
| Find groups in unlabeled data | Clustering |
| Simplify data with many features | Dimensionality reduction |
| Learn a sequence of actions from rewards | Reinforcement learning |
For the exam: Look at two things in the question: is the data labeled, and is the answer a category, a number, a group, or a sequence of actions?