What classification is

Classification predicts a category (class label) instead of a number.

Examples:

• spam vs not spam
• fraud vs normal
• cat vs dog
• churn vs not churn

Hands-On Machine Learning (Géron) introduces classification with the MNIST handwritten-digit dataset — 70,000 labeled images that are often called the “hello world” of Machine Learning. This phase follows that same chapter (3), plus the logistic/softmax half of Chapter 4, all of Chapter 5 (SVMs), and all of Chapter 6 (Decision Trees).

What you’ll learn

• how classification differs from regression, and why accuracy alone can lie to you
• Logistic Regression — sigmoid, log loss, and Softmax for multiple classes
• K-Nearest Neighbors — classifying by neighborhood vote
• Support Vector Machines — max-margin boundaries and the kernel trick
• Decision Trees — CART splits, Gini impurity vs entropy
• Naïve Bayes — a fast probabilistic baseline, especially for text
• how to evaluate with a confusion matrix, precision/recall/F1, and ROC/AUC

Phase 4 map

  flowchart TD
  A["Classification basics"] --> B["Logistic Regression"]
  B --> C["K-Nearest Neighbors"]
  C --> D["Support Vector Machines"]
  D --> E["Decision Trees"]
  E --> F["Naive Bayes"]
  F --> G["Confusion Matrix"]
  G --> H["Precision / Recall / F1"]
  H --> I["ROC & AUC"]

A note on metrics

Many classification datasets are imbalanced — think fraud detection, where maybe 1% of transactions are fraudulent. A classifier that always predicts “not fraud” can score over 99% accuracy while being completely useless. This is exactly the trap Géron walks through with a “never guess 5” classifier on MNIST: it hits over 90% accuracy while never detecting a single 5.

That’s why this phase spends real time on the confusion matrix, precision, recall, F1, and ROC/AUC — metrics built for exactly this kind of skewed, real-world data.