Heatmap Results
Formula Used
Topic normalization: P(term j | topic i) = x(i,j) / Σj x(i,j) Term normalization: P(topic i | term j) = x(i,j) / Σi x(i,j) Min-max scaling: x' = (x - xmin) / (xmax - xmin) Z-score scaling: z = (x - μ) / σ Cosine similarity: similarity(A,B) = (A · B) / (||A|| ||B||) Topic entropy: H(topic i) = -Σj p(i,j) log p(i,j)
How to Use
Paste a topic-term matrix or upload a supported file. Keep term names in the first row. Keep topic names in the first column.
Choose normalization, filtering, sorting, and appearance settings. Add a comparison matrix when needed. Generate the heatmap to inspect detailed results.
Click any heatmap cell for context. Review topic, term, similarity, and overlap tables. Export the chart and processed data afterward.
Example Data Format
| Topic | learning | model | data | network |
|---|---|---|---|---|
| Topic 1 | 0.34 | 0.29 | 0.18 | 0.05 |
| Topic 2 | 0.07 | 0.13 | 0.32 | 0.28 |
| Topic 3 | 0.11 | 0.09 | 0.14 | 0.36 |
Frequently Asked Questions
What does a topic-term heatmap show?
It displays term strength across discovered topics. Darker or brighter cells indicate stronger associations. Labels help interpret each topic.
Which topic models are supported?
The calculator accepts matrices from LDA, NMF, LSA, BERTopic, and custom workflows. It does not train models itself. It visualizes supplied weights.
Should topic rows sum to one?
Probability matrices commonly sum to one by row. Some model outputs use unnormalized weights. Row normalization can standardize those values.
Why use normalization?
Normalization makes topics or terms easier to compare. Different methods answer different analytical questions. Raw values remain available through exports.
How are dominant terms detected?
Terms are ranked by their processed topic weights. The highest values become dominant terms. The selected top-term count controls summaries.
What does topic entropy mean?
Entropy describes how broadly topic weight is distributed. Lower entropy suggests concentrated topics. Higher entropy indicates broader term mixtures.
How is topic similarity calculated?
You can select cosine, correlation, Euclidean, Manhattan, or Jensen-Shannon comparison. Similarity tables rank related topic pairs. Distance metrics are converted for clearer ranking.
Can two matrices be compared?
Yes, paste a matching comparison matrix. Choose difference or percentage-change mode. Labels and dimensions must align correctly.
Can large matrices be visualized?
Yes, but filtering improves readability and speed. Limit displayed terms and remove zero columns. Cell labels may be hidden automatically.