Three families of learning

Supervised learning is the workhorse, but it's one of three broad families, and knowing where a problem fits tells you which tools apply:

• Supervised — labeled examples; learn X to y. Spam detection, price prediction.
• Unsupervised — no labels; find structure in the data itself. Grouping customers, spotting anomalies.
• Reinforcement learning — learn by trial and error against a reward. Game-playing, robotics, some control and recommendation systems.

Most real business value today comes from the first two.

What unsupervised learning does

With no answer key, the model looks for patterns on its own:

• Clustering groups similar items — customer segments, related documents. (k-means is the classic.)
• Dimensionality reduction compresses many features into a few while keeping the signal — useful for visualization and as a preprocessing step.
• Anomaly detection flags the unusual — fraud, defects, outages.

You'll often combine them: cluster to discover segments, then build a supervised model per segment.

Classical ML vs. deep learning

Within all this sits another split. Classical ML — the focus of this course: linear models, trees, gradient boosting — excels on tabular data, the spreadsheets of rows and columns that run most businesses. It's fast, interpretable, and frequently wins on this kind of data. Deep learning dominates unstructured data — images, audio, language — where useful features are hard to hand-engineer.

The practical lesson: don't reach for a neural network because it's fashionable. For a table of customer records, a gradient-boosted tree is usually faster, cheaper, and more accurate.

Match the method to the data: classical ML for tables, deep learning for pixels and words, unsupervised when you have no labels. Most "which algorithm?" questions answer themselves once you name the data and the goal.

Try this: Think of a dataset from your world — transactions, messages, photos. Is it labeled or not? Tabular or unstructured? Those two questions alone narrow you from "all of ML" to a short, sensible shortlist of approaches.