Calculator settings
Formula used
The calculator builds a fuzzy neighbor graph before optimizing coordinates. Parameters control local scale, spacing, attraction, and repulsive movement. This educational implementation favors transparent calculations and moderate datasets.
How to use this calculator
Paste rows containing numeric features into the dataset field. Select header and label settings matching your table. Confirm every data row uses the same delimiter.
Choose a distance metric matching your feature meaning. Set neighbors higher for broader structures and smoother regions. Lower minimum distance creates tighter visual groups.
Run the calculator and inspect preservation metrics first. Compare several seeds before trusting one visual arrangement. Download coordinates for reporting or downstream cluster analysis.
Understanding UMAP projections
Each row represents one observation with several measured features. Distances summarize similarity before any projection coordinates exist. The chosen metric therefore shapes every neighborhood relationship.
Preparing reliable input data
Clean missing values before submitting data to this calculator. Numeric columns should describe comparable analytical concepts whenever possible. Remove identifiers unless they provide meaningful distance information.
Choosing neighborhood size
Start near five neighbors for small demonstration datasets. Increase gradually while watching trustworthiness and neighborhood preservation. Strong conclusions should remain visible across reasonable parameter changes.
Controlling compactness and spread
Spread controls the overall scale of projected coordinates. It works together with minimum distance during similarity modeling. Compare both settings instead of changing either parameter blindly.
Reading quality diagnostics
Trustworthiness penalizes unexpected neighbors appearing after dimensionality reduction. Continuity checks whether original neighbors remain visible after projection. Both values approach one when neighborhoods remain faithful.
Neighbor preservation reports direct overlap between neighborhood lists. Stress summarizes normalized distance distortion across all observation pairs. Silhouette uses labels and only appears when groups repeat.
Interpreting the finished graph
Nearby points usually represent observations with similar feature profiles. Empty gaps may indicate separation, scaling effects, or optimization artifacts. Axes have no direct original-feature meaning.
Colors display supplied labels without influencing coordinate optimization. Visible edges expose strong fuzzy graph relationships between observations. Dense edge bundles suggest robust local connections.
Improving analytical confidence
Repeat calculations using several seeds and neighborhood values. Stable patterns deserve more confidence than isolated visual arrangements. Check suspicious clusters against original features and domain knowledge.
UMAP supports exploration rather than automatic causal interpretation. Projection geometry can exaggerate gaps or compress meaningful variation. Validate every finding with additional statistical methods.
Performance and reporting
This page limits rows to protect shared hosting performance. Pairwise distances grow quickly as dataset size increases. Sampling can preserve broad structure for initial investigation.
Download coordinates when another workflow needs projected features. Record every parameter beside exported results for reproducibility. Careful documentation makes comparisons more useful and defensible.
Practical modeling guidance
Use cosine distance for directional similarity between feature vectors. Manhattan distance may suit sparse or robust comparisons. Euclidean distance remains useful for standardized continuous measurements.
Explore multiple settings before selecting a presentation graphic. Report diagnostics alongside plots, labels, and preprocessing decisions. Transparent choices strengthen every conclusion drawn from visual embeddings.
Final review
Confirm labels match rows before interpreting colored clusters. Review outliers with original values and collection context. Reliable embeddings require thoughtful preparation, testing, and cautious interpretation.
Use the graph as one analytical perspective only. Combine it with metrics, models, and subject expertise. Strong evidence comes from repeated checks across several methods.
Frequently asked questions
1. What does this UMAP calculator produce?
It produces two-dimensional or three-dimensional coordinates from numeric feature rows. The page also reports trustworthiness, continuity, neighborhood preservation, stress, and optional silhouette values.
2. Which distance metric should I choose?
Use Euclidean distance for standardized continuous features. Choose Manhattan distance for robust absolute differences. Cosine distance works well when direction matters more than magnitude.
3. How should I select nearest neighbors?
Smaller values emphasize local detail and uncommon groups. Larger values reveal broader continuity and global organization. Test several values and compare the reported preservation metrics.
4. What does minimum distance change?
Minimum distance controls compactness within the final embedding. Lower values allow tighter clusters. Higher values preserve more spacing among nearby observations.
5. Why do different seeds change the plot?
Optimization begins from chosen starting coordinates and follows nonlinear updates. Different seeds can produce rotated, mirrored, or locally rearranged results. Stable patterns should persist across several runs.
6. Should features always be standardized?
Standardization is recommended when columns use different units or scales. It prevents large numeric ranges from controlling distance calculations. Keep original scales only when those differences are meaningful.
7. Can I use text or categories directly?
No. Feature columns must contain numeric values. Encode categories first. Convert text into numeric vectors using suitable feature extraction methods.
8. What indicates a reliable projection?
Higher trustworthiness, continuity, and neighbor preservation indicate stronger local fidelity. Lower stress indicates better distance agreement. Reliability also requires stability across parameters and seeds.
9. Can the coordinates train another model?
They can support experiments, visualization, or downstream features. However, transformations should be fitted consistently for production data. Validate performance carefully before using coordinates in operational models.