Embedding Results
Results appear here after calculation.
Embedding Coordinates
| Run a calculation |
|---|
| No coordinates available. |
Run Summary
Selected Point
Processing Log
Formula Used
t-SNE converts distances into neighborhood probabilities. It compares high-dimensional and embedded similarities. Optimization minimizes their Kullback-Leibler divergence.
Perplexity controls the effective neighborhood size. Early exaggeration separates groups during initial optimization. Student probabilities reduce crowding in the map.
How to Use
- Paste data or load a sample dataset.
- Preview columns and select numeric features.
- Choose preprocessing and t-SNE parameters.
- Run one or several seeded embeddings.
- Inspect clusters, neighbors, and diagnostics.
- Export coordinates, images, or reports.
Example Data
| id | label | feature_1 | feature_2 | feature_3 |
|---|---|---|---|---|
| item-1 | A | 1.20 | 0.85 | 2.40 |
| item-2 | A | 1.05 | 0.91 | 2.18 |
| item-3 | B | 4.70 | 3.95 | 0.55 |
Interpretation Guide
Nearby points often share local structure. Large map distances can be misleading. Axes have no direct feature meaning.
Cluster shapes depend on parameters and seeds. Compare several runs before drawing conclusions. t-SNE does not prove clusters exist.
Frequently Asked Questions
What perplexity should I use?
Start near thirty for moderate datasets. Test smaller and larger values. Stable patterns deserve greater confidence.
Why do runs look different?
Random initialization affects optimization. Fix the seed for reproducibility. Compare multiple seeds for stability.
Should features be scaled?
Usually yes for mixed numeric ranges. Scaling prevents large units dominating distances. Check domain meaning before transformation.
What does KL divergence mean?
It measures neighborhood mismatch. Lower values indicate better objective fitting. Values across datasets remain incomparable.
Can t-SNE detect clusters?
It visualizes local relationships. It can suggest grouping patterns. Use clustering methods for formal assignments.
Why use PCA first?
PCA removes noise and redundancy. It can improve speed substantially. Retain enough meaningful variation.
Are axis values meaningful?
Their orientation is arbitrary. Rotations preserve the same structure. Interpret neighborhoods instead of coordinates.
How many iterations are enough?
One thousand often provides a baseline. Hard datasets may need more. Watch the objective curve.
Can I use categorical features?
Encode them before calculation. One-hot encoding is commonly suitable. Choose distances matching data semantics.