COMP 2201 · Machine Learning: Data Preprocessing

Led by Fisherian Statistical Learning Simulacrum

About this course

Fisher's principle: the answer must not be used to design the question. In machine learning, this means the test set must never influence any preprocessing decision — not the imputation mean, not the scaling parameters, nothing. Every number you compute from training data must be applied to test data, never the reverse.

This course covers the complete preprocessing pipeline in Python with scikit-learn: importing data, handling missing values, encoding categorical features, splitting into training and test sets, and scaling features. Every step is grounded in the statistical principle it expresses.

Based on the foundational writings of Ronald Fisher.

1. Module 1

   The Machine Learning Workflow

   Led by Fisherian Statistical Learning Simulacrum

   The question

   Data import, preprocessing, model training, evaluation, deployment — what is the purpose of each stage, and what mistake is possible at each? Why must training and test data be kept strictly separate from the first line of code?

   Outcome

   The student can describe the ML workflow and import a dataset identifying features and labels.

   Sub-units

   1. ○ 1.1 The Workflow Open this unit → ✓ ↻
   2. ○ 1.2 Import a Dataset Open this unit → ✓ ↻
   Begin the course → Print certificate of completion
2. Module 2

   Handling Missing Data

   Led by Fisherian Statistical Learning Simulacrum

   The question

   Mean imputation replaces missing values with the column average. This is wrong in many situations and appropriate in others. What determines which strategy is justified — and why must the imputer be fit on training data only?

   Outcome

   The student can apply SimpleImputer and explain when mean imputation is appropriate.

   Sub-units

   1. ○ 2.1 Detect and Impute Open this unit → ✓ ↻
   2. ○ 2.2 Essay: Strategy Choice Open this unit → ✓ ↻
   Open this module → Print certificate of completion
3. Module 3

   Encoding Categorical Data

   Led by Fisherian Statistical Learning Simulacrum

   The question

   Assign 0, 1, 2 to three countries and you have implied an ordering that does not exist. One-hot encoding avoids this — but creates the dummy variable trap. What is the trap, and how does dropping one column avoid it?

   Outcome

   The student can apply one-hot encoding and explain the dummy variable trap.

   Sub-units

   1. ○ 3.1 One-Hot Encode Open this unit → ✓ ↻
   Open this module → Print certificate of completion
4. Module 4

   Training and Test Split

   Led by Fisherian Statistical Learning Simulacrum

   The question

   If you compute the imputation mean on the full dataset (training + test), you have contaminated your evaluation. What exactly goes wrong — and what is the correct order of operations?

   Outcome

   The student can apply train_test_split and explain the contamination principle.

   Sub-units

   1. ○ 4.1 Split the Data Open this unit → ✓ ↻
   2. ○ 4.2 The Contamination Question Open this unit → ✓ ↻
   Open this module → Print certificate of completion
5. Module 5

   Feature Scaling

   Led by Fisherian Statistical Learning Simulacrum

   The question

   A salary column with values 0-100,000 and a years-of-experience column with values 0-10 will mislead any distance-based algorithm. Standardisation corrects this. Which algorithms require it — and why must the scaler be fit on training data only?

   Outcome

   The student can apply StandardScaler correctly and describe a complete preprocessing pipeline.

   Sub-units

   1. ○ 5.1 Apply StandardScaler Open this unit → ✓ ↻
   2. ○ 5.2 Final Essay: The Complete Preprocessing Pipeline Open this unit → ✓ ↻
   Open this module → Print certificate of completion