A short, focused unit that doubles as the second-quarter capstone. The central problem: credit card fraud is rare — less than 0.2% of transactions — so a classifier that always predicts “not fraud” gets 99.8% accuracy and catches nothing. This unit explores how to handle severe class imbalance and how to detect anomalies without a clean supervised signal.

Concepts You’ll Learn About

  • Class imbalance — why accuracy is a misleading metric when one class dominates; precision, recall, and F1 revisited
  • SMOTE — Synthetic Minority Oversampling Technique; generating synthetic examples of the rare class
  • Gaussian Mixture Models — unsupervised density estimation; flagging points in low-probability regions as anomalies
  • Supervised vs. unsupervised anomaly detection — when you have labeled fraud examples vs. when you don’t

Topics

  • Credit card fraud — baseline and SMOTE — the raw imbalance problem, then oversampling the fraud class and measuring the impact on recall.

  • Gaussian Mixture Model approach — fitting a GMM to the transaction data and flagging low-likelihood points as anomalies; comparing to the supervised approach.

Quarter Project

Find any dataset from the UCI Repository and model it using techniques from Units 1–8. Classification, regression, or anomaly detection — your choice. The deliverable is a notebook with a clear question, a defensible modeling approach, and a written interpretation of results.

What’s next

Unit 09 introduces dense neural networks — perceptrons, backpropagation, and the foundational ideas that make everything in Units 10–13 possible.