Hierarchical Clustering Heatmap

Use this page as a cautious guide to dendrogram-style hierarchical grouping, not as a claim that the current product exposes a full clustered heatmap.

Figure purpose

This page should be read cautiously. The current implementation is better described as a hierarchical-clustering dendrogram view than as a full clustered heatmap with integrated matrix display.

That still makes it useful for exploratory grouping. It can help show how observations or items merge under one current hierarchical procedure.

What the figure shows

The current public surface should be understood as a dendrogram-first clustering view. In practice, the output is meant to show:

  • The relative merge order of observations or features under one hierarchical clustering procedure
  • Branch structure that helps you see which items join early versus late
  • Relative closeness in the sense encoded by the current clustering pipeline, rather than a claim of one true underlying grouping

The page should not be read as promising a full matrix heatmap with fully integrated row-and-column clustered display unless that broader contract is explicitly surfaced elsewhere in the product.

When to use or avoid

Use this page when the main question is how items join together in one hierarchical grouping view. It can help readers inspect relative closeness and merge order in an exploratory way.

Avoid treating the figure as proof that the discovered grouping is the true structure of the data. Linkage choice, distance assumptions, preprocessing, and the current implementation boundary all matter.

Because the page title is broader than the current support surface, keep the wording dendrogram-first.

Licklider cannot determine automatically which distance metric, linkage strategy, preprocessing path, or cluster interpretation is scientifically correct for your question. It also does not tell you how many clusters are "really" present or whether a visually appealing branch split is stable enough to treat as a substantive finding.

These limits matter because hierarchical clustering can look authoritative even when small preprocessing or scaling changes would alter the merge order. The figure is useful for pattern finding, but it is not a proof that the displayed tree is the one correct structure in the data.

Required columns

  • Multiple numeric values that can support hierarchical grouping
  • An exploratory reading task where the merge structure itself is of interest

The current public page should not imply that a full heatmap-plus-dendrogram contract is already available.

Hierarchical clustering is exploratory. It helps organize one view of similarity or grouping, but it does not automatically discover the one correct clustering for a dataset.

The current implementation boundary is also important. This page should not be written as if row/column clustered heatmap behavior is fully exposed in the same surface.

Design Rationale & References

This page follows a simple rule: exploratory clustering views should show one concrete grouping procedure without overstating that procedure as a definitive structure discovery tool. That is why the current documentation is written dendrogram-first and avoids promising a full clustered-heatmap suite when the exposed surface is narrower.

The caution around interpretation is intentional. Hierarchical clustering depends on modeling choices such as distance definition, linkage rule, scaling, and preprocessing. Those choices can materially change merge order and apparent group boundaries, so the figure is framed as an exploratory similarity view rather than a claim that the product has identified the one correct set of clusters [1, 2].

The page also distinguishes this view from a standard heatmap on purpose. A heatmap emphasizes matrix pattern display, whereas a dendrogram emphasizes merge order and relative grouping. Keeping that distinction visible helps readers pick the surface that matches their question instead of treating all multivariate plots as interchangeable.

  1. Ward, J. H. Jr. (1963). Hierarchical Grouping to Optimize an Objective Function. Journal of the American Statistical Association, 58(301), 236-244. https://doi.org/10.1080/01621459.1963.10500845
  2. Sokal, R. R., & Michener, C. D. (1958). A Statistical Method for Evaluating Systematic Relationships. University of Kansas Science Bulletin, 38, 1409-1438.

Alternative figures

  • Use Heatmap when the matrix display itself is the main priority.
  • Use K-means Clustering Plot when you want to inspect one partition-style clustering overlay instead of a dendrogram.
  • Use Parallel Coordinates when per-observation multi-axis patterns matter more than tree structure.

TODO (Phase02+)

  • Expand only if the public product surface later supports a clearer combined heatmap-plus-dendrogram contract.