Week 13 Guide

Chapter 3: Dimensionality Reduction Application

Week 12 introduced dimensionality reduction and PCA from sklearn.decomposition. You saw how correlated features create redundancy in a dataset, how PCA finds directions of maximum variance and re-expresses samples in terms of those directions, why scaling is required before PCA, how the fit-on-training rule applies to the PCA object exactly as it applies to scalers, and how to interpret explained_variance_ratio_ when choosing how many components to retain. This week you apply those concepts in a new context.

The assignment puts you in the role of a consultant advising a data scientist at a game studio. Each task draws directly on what you built in Week 12.

Week 13 Assignment:

• Here is the link to the Week 13 assignments page.

What you need to be ready to apply

Why correlated features motivate dimensionality reduction. You should be able to look at a set of features, identify which ones are likely to overlap in what they measure, and explain why that redundancy is a problem worth addressing before training. You should also be able to distinguish this clearly from the scale problem that scaling addresses. These are two different problems with two different solutions.

The fit-on-training rule applied to PCA. You applied this rule to scalers in Week 10. Here it applies to a second object in the preprocessing sequence. You should be able to explain what PCA learns during fit, how that is different from what a scaler learns, and what goes wrong when the rule is violated for either object.

Reading explained_variance_ratio_ and choosing n_components. You should be able to interpret a full explained variance table, reason about where meaningful variance ends and noise begins, and make a defensible recommendation for how many components to retain. You should be able to explain the tradeoff you are accepting and push back on a recommendation that does not hold up to scrutiny.

Why fewer components can outperform more features, and what comes next. You should be able to explain the mechanism behind cases where a kNN model on PCA-reduced data outperforms a model on all scaled features. You should also know what t-SNE offers for visualization and why it cannot serve as a preprocessing step, and what inverse_transform does when a data scientist needs to understand what the components represent.

Week 13 tasks

1. Review your Week 12 demo notebook and textbook reading (Chapter 3, pages 140–165) as needed before working through the assignment.
2. Read through the Week 13 assignment carefully, including the full scenario and all four sections, before recording.
3. Record and submit your video addressing all four sections.