COMP 2203 · Machine Learning: Classification

Led by Pearlian Causality Simulacrum

Seven classifiers — from Bayes's posterior probability to Random Forest ensembles — with Python implementation and systematic evaluation.

1. Module 1

   Logistic Regression and Classification Fundamentals

   Led by Pearlian Causality Simulacrum

   The question

   The sigmoid function maps any linear combination to a probability between 0 and 1. Logistic regression uses this to estimate the probability of class membership directly. What does the decision boundary look like — and what does the confusion matrix reveal that accuracy hides?

   Outcome

   The student can implement logistic regression, produce a confusion matrix, and visualise the decision boundary.

   Sub-units

   1. ○ 1.1 Logistic Regression Implementation Open this unit → ✓ ↻
   2. ○ 1.2 Visualise the Decision Boundary Open this unit → ✓ ↻
   Begin the course → Print certificate of completion
2. Module 2

   K-Nearest Neighbours and SVM

   Led by Pearlian Causality Simulacrum

   The question

   KNN predicts by majority vote among the k nearest training points — all computation at prediction time. SVM finds the hyperplane that maximises the margin between classes. When do these simple geometric classifiers fail — and what is the same dataset telling you when one outperforms the other?

   Outcome

   The student can implement KNN and SVM, explain margin maximisation, and identify when linear classifiers fail.

   Sub-units

   1. ○ 2.1 KNN and SVM Open this unit → ✓ ↻
   Open this module → Print certificate of completion
3. Module 3

   Kernel SVM and Naive Bayes

   Led by Pearlian Causality Simulacrum

   The question

   The RBF kernel maps data into an infinite-dimensional space where it becomes linearly separable, then draws the boundary back in the original space. Naive Bayes applies Bayes's theorem assuming features are independent — which is almost always false. Why does it still work?

   Outcome

   The student can implement kernel SVM and Naive Bayes and explain their respective assumptions.

   Sub-units

   1. ○ 3.1 Kernel SVM and Naive Bayes Open this unit → ✓ ↻
   Open this module → Print certificate of completion
4. Module 4

   Decision Tree and Random Forest Classification

   Led by Pearlian Causality Simulacrum

   The question

   A random forest trains each tree on a different bootstrap sample AND a different random subset of features. Why does double randomness produce better generalisation — and what can feature importances tell you about the problem?

   Outcome

   The student can implement both, compare decision boundaries, and extract feature importances.

   Sub-units

   1. ○ 4.1 Decision Tree and Random Forest Open this unit → ✓ ↻
   Open this module → Print certificate of completion
5. Module 5

   Model Evaluation and Classifier Selection

   Led by Pearlian Causality Simulacrum

   The question

   A tool that predicts "negative" for every case in a 1% prevalence dataset has 99% accuracy. Why is this useless — and which metric should guide a medical screening classifier where false negatives cost lives?

   Outcome

   The student can compute precision, recall, and F1 and justify a classifier selection for a specific context.

   Sub-units

   1. ○ 5.1 Final Essay: Classifier Selection Open this unit → ✓ ↻
   Open this module → Print certificate of completion