PCA
- Form
- Linear.
- Score
- Variance.
- Components
- A fixed number of components.
Loading…
PCA keeps the directions of greatest variance, so the two fitted components are a linear summary of the columns in that table. t-SNE places each row so that its neighbors still stay neighbors, and that map follows the local neighborhoods.
Choose the map the table needs.
This is a choice, not a measured winner.
A new row uses the fitted PCA axes; t-SNE has no transform, and this product places a later row by 5-neighbor interpolation into the fitted map, which is approximate.
t-SNE keeps more original neighborhoods on Iris Flowers, Wine (UCI wine recognition), and Customer Segments; PCA still reports the variance those two axes keep: 95.8% on Iris Flowers, 55.4% on Wine (UCI wine recognition), 90.5% on Customer Segments.
Do you need the directions of greatest variance?
A linear summary of the columns, or a map where neighbors stay neighbors.
PCA on Iris Flowers (150 rows, 4 features) keeps 95.8% of the variance in two linear components, trustworthiness 0.974. t-SNE with perplexity 30 and a PCA start has trustworthiness 0.991 and KL divergence 0.170. Both scatters are colored by held-out species, which neither fit receives. t-SNE keeps more of the original neighborhoods in this picture.
On Iris, PCA reports 95.8% of the variance in 2 components, and the t-SNE artifact records no explained variance.


The link loads the sample. Run is still required.
PCA on Wine (UCI wine recognition) (178 rows, 13 features) keeps 55.4% of the variance in two linear components, trustworthiness 0.871. t-SNE with perplexity 30 and a PCA start has trustworthiness 0.962 and KL divergence 0.367. Both scatters are colored by held-out wine_class, which neither fit receives. t-SNE keeps more of the original neighborhoods in this picture.


The link loads the sample. Run is still required.
PCA on Customer Segments (200 rows, 5 features) keeps 90.5% of the variance in two linear components, trustworthiness 0.927. t-SNE with perplexity 30 and a PCA start has trustworthiness 0.979 and KL divergence 0.343. Both scatters are colored by held-out segment, which neither fit receives. t-SNE keeps more of the original neighborhoods in this picture.


The link loads the sample. Run is still required.
Blob size, tightness, and gaps in the t-SNE picture are layout; perplexity 30 sets the neighborhood.
| Dataset | Method | Trustworthiness | Also reported |
|---|---|---|---|
| Iris Flowers | PCA | 0.974 | 95.8% variance in 2 components |
| Iris Flowers | t-SNE | 0.991 | KL divergence 0.170 |
| Wine (UCI wine recognition) | PCA | 0.871 | 55.4% variance in 2 components |
| Wine (UCI wine recognition) | t-SNE | 0.962 | KL divergence 0.367 |
| Customer Segments | PCA | 0.927 | 90.5% variance in 2 components |
| Customer Segments | t-SNE | 0.979 | KL divergence 0.343 |
PCA keeps the directions of greatest variance, so the two fitted components are a linear summary of the columns in that table. t-SNE places each row so that its neighbors still stay neighbors, and that map follows the local neighborhoods.
This is a choice, not a measured winner.
On Iris, PCA reports 95.8% of the variance in 2 components, and the t-SNE artifact records no explained variance.
t-SNE does not rank or drop columns; PCA loadings do. These compare runs do not publish loadings.
Method: StandardScaler, drop constant columns. Comparison code version 8fbe69461da4 (scripts/seo/build_compare_runs.py). scikit-learn 1.4.2. Seed 42. The class column was held out of every fit. Iris Flowers: UCI Machine Learning Repository, Iris (Fisher, 1936). License: CC BY 4.0. Wine (UCI wine recognition): UCI Machine Learning Repository, wine recognition. License: CC BY 4.0. Customer Segments: Synthetic customer records bundled with this site. License: Generated for demos; no external license.
Results computed, published, and last updated on 2026-11-03.
The comparison sentences are filled from these published runs by a fixed template, so the scores can be checked against the same datasets in the tool.
Maintainer: nikolaou1982.
Contact: nick1982100@gmail.com.