Interview: 'you have 500 features, most correlated — how do you preprocess?'
mediumAnswer
- (1) Correlation heatmap → drop obviously redundant.
- (2) Domain-driven grouping (aggregate related features first).
- (3) Missing-value pattern check.
- (4) Standardize.
- (5) PCA / Truncated SVD → keep 95% variance (usually 30-80 components).
- (6) If non-linear structure suspected: UMAP (for viz), autoencoder (for downstream).
- (7) For interpretable model: keep top-loadings; for pure prediction: use the full transformation.
- (8) Consider L1-regularized model as alternative that learns sparsity end-to-end.
Check yourself — multiple choice
- Random
- Correlation prune + domain grouping + missing check + standardize + PCA/TruncSVD 95% + UMAP/AE for non-linear + L1 as sparse alternative
- Only PCA
- Not real
500 correlated features: prune + standardize + PCA + non-linear if needed + L1 alt.
#interview#dimensionality-reduction
Practise Unsupervised Learning
214 interview questions in this topic.
Related questions
- Interview: 'when should you use PCA vs autoencoder for dim reduction?'
- PCA vs t-SNE vs UMAP — when do you use each?
- How do you decide how many PCA components to keep?
- Why does clustering degrade in high dimensions?
- Interview: 'you're asked to segment 10M customers — how?'
- Derive PCA — what does it optimize?