Heatmap Inputs and Clustering Options
Paste a matrix, select preprocessing, then build the clustered view.
Understanding Cluster Heatmaps
Cluster heatmaps combine matrix intensity with rows and columns. Similar observations move together, revealing groups hidden within raw tables. Colors then expose gradients, blocks, boundaries, and unusual local values.
The calculator accepts rectangular matrices containing features or measured responses. Row labels can identify samples, customers, documents, experiments, or models. Column labels can describe variables, dimensions, periods, or learned components.
Scaling and Normalization
Scaling changes how each column contributes to visible heatmap structure. Z-score scaling centers values and divides by standard deviation. Min-max scaling maps each column between zero and one.
Formula Used
Distance metrics define similarity before hierarchical clusters are assembled. Euclidean distance measures straight-line separation across every feature. Manhattan distance totals absolute differences and often handles spikes better.
Cosine distance compares vector directions rather than their numeric magnitudes. It works well for sparse embeddings and frequency-based representations. Choose metrics that match the meaning behind your data.
Linkage rules control distances between growing hierarchical cluster groups. Single linkage uses the nearest pair across two candidate groups. Complete linkage uses the farthest pair and favors compact clusters.
Average linkage measures every cross-group pair before calculating mean separation. It often balances chaining behavior against overly tight cluster shapes. Different linkage choices can produce meaningfully different row orders.
Missing Values and Ordering
Row clustering groups observations using all currently visible matrix columns. Column clustering groups features using values across the matrix rows. Enable either direction separately when domain ordering must remain fixed.
Missing entries can be rejected, replaced, or interpreted as zeros. Mean replacement preserves column centers but may reduce visible variability. Median replacement can resist distortion from skewed measurements.
How to Use This Calculator
The result summary reports matrix size, range, mean, and variation. Ordered labels help connect visible blocks with original entities. The interactive plot supports hovering, zooming, downloading, and precise inspection.
Value annotations improve small matrices but can clutter dense displays. Decimal controls keep annotations readable and consistent across cells. Transposition switches analytical orientation without changing the original input.
Reading the Result
Cluster heatmaps assist exploratory analysis before modeling and feature engineering. They can expose duplicated samples, correlated variables, and suspicious batches. They also support segmentation reviews and representation quality checks.
Interpret strong blocks as hypotheses rather than guaranteed statistical conclusions. Scaling, metric choice, and linkage can change apparent structures. Validate discovered groups with domain evidence and independent measurements.
Use consistent preprocessing when comparing heatmaps across experiments or datasets. Record selected options beside exported figures for reproducible analysis. Stable settings make visual comparisons clearer for future decisions.
Euclidean Distance
d(x,y) = √Σ(xᵢ − yᵢ)²This measures direct geometric separation between two vectors.
Z-Score Scaling
z = (x − μ) / σThis centers each feature and adjusts its spread.
Average Linkage
D(A,B) = Σ d(a,b) / (|A||B|)This averages every pairwise distance across two clusters.
Practical Steps
- Paste a rectangular numeric matrix into the main field.
- Add optional row and column labels for interpretation.
- Select scaling, distance, linkage, and missing-value rules.
- Choose clustering directions, colors, clipping, and annotations.
- Submit the form and inspect the reordered heatmap.
Frequently Asked Questions
1. Which matrix formats are accepted?
Use one observation per line. Separate values with commas, spaces, semicolons, or tabs. Every row must contain equal columns.
2. Are row and column labels required?
Labels are optional. Missing names receive automatic row or column numbers. Clear labels improve cluster interpretation.
3. When should Z-score scaling be selected?
Use Z-scores when features use different units. Each feature receives comparable influence. Constant columns become zero.
4. Which distance metric is best?
Euclidean distance suits continuous standardized features. Manhattan distance tolerates isolated jumps. Cosine distance emphasizes direction.
5. How does linkage affect clusters?
Single linkage can create long chains. Complete linkage favors compact groups. Average linkage provides a balanced compromise.
6. How are missing values handled?
You can reject or replace missing entries. Mean and median choices use column statistics. Zero replacement may alter distances.
7. What does transposing the matrix change?
Transposition swaps observations and features. Existing labels swap with their matching dimensions. Clustering then uses the new orientation.
8. Why can large matrices feel slower?
Hierarchical clustering compares many vector pairs. Larger dimensions require more repeated distance calculations. Reduce matrix size when needed.
9. Can the heatmap be exported?
Use the chart toolbar for image export. Hovering reveals exact labels and values. Saved figures preserve visual results for later reporting.