AI & ML: Supervised vs Unsupervised Learning

The Two Main Families

Almost every machine learning problem falls into one of two families, distinguished by whether the data has labels. Supervised learning trains on examples that include the answer. Unsupervised learning finds structure in data that has no answer column at all. A third family, reinforcement learning, learns by trial and error with rewards, and powers game playing and robotics.

Supervised Learning

Labelled data is a collection of rows where each row has features and a known outcome. The model learns to predict the outcome from the features.

# labelled rows: (features) -> label
(150kg, 165cm) -> high risk
(70kg, 175cm)  -> low risk
(95kg, 180cm)  -> medium risk

Classification predicts a category such as high risk or low risk. Regression predicts a number such as a price or temperature. Both live under the supervised umbrella because both need labelled examples.

Unsupervised Learning

Unsupervised data has no answers to check against. The algorithms instead look for patterns: grouping similar rows into clusters, finding unusual rows, or compressing many dimensions into few.

• Clustering - group customers into segments with similar behaviour.
• Anomaly detection - flag transactions that look unlike the rest.
• Dimensionality reduction - shrink thousands of features to a few.

An online store might cluster shoppers by purchase history without anyone telling the algorithm which customers belong together. The label is discovered, not provided.

Choosing an Approach

if labelled data and predict a category:  supervised classification
if labelled data and predict a number:   supervised regression
if no labels and find groups:            unsupervised clustering
if no labels and spot outliers:          unsupervised anomaly detection

Labels are expensive because humans usually create them, so unsupervised methods are attractive when data is abundant but answers are not. In practice, teams often mix both: cluster first, then label just the interesting clusters.

Key Points

• Supervised learning needs labelled examples to map features to answers.
• Classification predicts categories and regression predicts numbers.
• Unsupervised learning finds clusters and outliers without labels.
• Reinforcement learning improves through rewards over many tries.