2. Build a First Pipeline

Introduction

You will now build the smallest complete Machine Learning pipeline: load, clean, split, engineer features, train, evaluate. Every step matters, but the most important one is the split: instead of shuffling all rows together, you will train on one city and test on another. A model that only works on the city it was trained on is not a model you can trust anywhere else.

Notebook

Files: notebooks/session2/02_pipeline.ipynb

Open it and select the same .venv kernel you used in the Notebook section of the previous activity. Complete each # TODO cell in order.

Step 1 — Load and restrict the scope

Load data/train.csv, keep only Kampala and Nairobi and the columns listed in COLUMNS, then cap the sample at 1200 rows per city (random, random_state=42) so the pipeline stays fast to run and easy to inspect.

Step 2 — Clean missing values

Some satellite passes are missing for some columns. Fill them within each city, sorted by date (forward-fill then backward-fill). Never let one city’s values fill another city’s gaps — that would be a geographic data leak before you even reach the model.

Step 3 — Manual split by city

No random shuffling of rows. train_df is one city, test_df is the other. Testing on data the model never trained on is the basic way to check that it actually generalizes. Later, in Session 3, you will see a more systematic way to do this across several groups at once — cross-validation.

Step 4 — Temporal features

Extract month and dayofweek from date. Pollution often follows seasonal and weekly patterns (traffic, economic activity).

Step 5 — Baseline model and evaluation

Train a plain LinearRegression and evaluate it with RMSE, MAE and R². The notebook explains why site_latitude/site_longitude are deliberately left out of the features — read it before you try adding them back.

Reflection

Files: reflection/session_2/module_2.md