PCA Transformation and Reconstruction
You'll learn: How to project data into principal component space and reconstruct it back, understanding what information gets lost in the process.
The Core Idea
Once you've identified your principal components through eigendecomposition, PCA performs two key operations:
- Transformation (projection): Converting original high-dimensional data into the lower-dimensional PC space
- Reconstruction: Mapping the compressed representation back to the original space
Think of it like taking a 3D sculpture and photographing it from the "best angle" (the principal components). You can display that 2D photo, but if you try to recreate the sculpture from just the photo, you'll miss some details—that's your information loss.
How It Works
Transformation multiplies your centered data by the principal component matrix. If you keep only k components instead of all p, you're compressing the data.
Reconstruction reverses this by multiplying the transformed data back using the transpose of the PC matrix. However, if you dropped components, the reconstructed data won't perfectly match the original—it's an approximation.
The difference between original and reconstructed data is the approximation error (also called reconstruction error). This error comes from the variance you discarded by not keeping all components.
Why This Matters
Transformation gives you compressed features for modeling. Reconstruction tells you how much information you sacrificed. The variance explained by your chosen components (from lesson 389) directly relates to reconstruction quality: more variance retained = smaller reconstruction error.
Key Takeaway: PCA transformation projects data into PC space for compression; reconstruction maps it back with some information loss proportional to the discarded variance.