QWAV · Interactive Artifact A2

Q-PNA Classifier Playground

A glass-box classifier. Pick a hierarchical dataset, watch agglomerative clustering build an ultrametric tree in real time, read the cophenetic loss, and click any leaf to see — in plain steps — why it was classified where it was.

GLASS-BOX · tree embedding
Dendrogram · ultrametric tree click a leaf to trace its decision path
merge height = cophenetic distance
Cophenetic loss (1 − CPCC)
tree fidelity vs. raw distances
CPCC
cophenetic correlation coefficient
Clusters found
at default cut height
Neural baseline accuracy
linear probe trained on same labels

Tree embedding (LinMap view) — merge heights as coordinates

Glass-box: why was this point classified here?

Click a leaf in the dendrogram to trace its full decision path — the sequence of merges that determined its class.

What Q-PNA-style classification does

The classifier builds a hierarchical tree by repeatedly merging the two closest clusters (agglomerative clustering). The height at which two points first merge is their cophenetic distance — the tree's internal measure of similarity. A high CPCC (close to 1) means the tree faithfully preserves the original geometry.

CPCC = Pearson corr( d_ij , c_ij )

The LinMap view on the right embeds each point at the merge heights along its path to the root — an ultrametric (tree) coordinate system. This is the "glass box": every classification has a visible, inspectable structure, unlike a weight matrix.

Tree vs. neural baseline

Press Train Neural Baseline to fit a small linear classifier on the same labeled points. It reaches an accuracy score but gives you only probabilities — no reason, no path, no hierarchy.

The tree gives an interpretable merge history for every leaf. That is the glass-box property the Q-PNA approach is designed to expose: why a point is grouped with its neighbors, step by step.

† Pedagogical model: demonstrates interpretable hierarchical (ultrametric) classification and compares it with a simple neural baseline. It is not a claim about proprietary Q-PNA software or physical hardware. See QWAV Research Specification (Q-PNA v2.0) for the underlying method.