cddiagram

A pure Python library for generating Critical Difference (CD) diagrams as SVG.

CD diagrams visualize the statistical comparison of multiple classifiers (or models) over multiple datasets, as introduced by Demsar (2006). They show the average rank of each model and connect groups of models whose performance differences are not statistically significant.

J. Demsar, "Statistical Comparisons of Classifiers over Multiple Data Sets", Journal of Machine Learning Research, vol. 7, pp. 1-30, 2006. https://jmlr.org/papers/v7/demsar06a.html

How it works

1. A Friedman test checks whether at least one model differs significantly from the others (at alpha = 0.05).
2. If significant, a pairwise post-hoc test determines which pairs of models differ. Two options are supported:
   • posthoc="wilcoxon-holm" (default): pairwise two-sided Wilcoxon signed-rank tests with Holm step-down correction. Each pairwise decision uses only those two models' per-dataset scores, so it is robust to which other models are in the experiment (Benavoli, Corani & Mangili, 2016).
   • posthoc="nemenyi": original Nemenyi test from Demšar (2006). A single critical distance CD = q_α · √(k(k+1)/6N) is computed; pairs with |R̄_i − R̄_j| ≤ CD are deemed non-significant. The diagram includes the iconic CD bar.
3. Maximal contiguous cliques of mutually non-significant models are connected by horizontal bars below the rank axis.
4. The result is rendered as an SVG diagram showing ranked models and significance groups.

Install

pip install cddiagram

Requires Python 3.12+ and depends on numpy and scipy.

Usage

Write to file

import numpy as np
from cddiagram import draw_cd_diagram

rng = np.random.default_rng(1)

models = {
    "model1": rng.normal(loc=0.2, scale=0.1, size=30),
    "model2": rng.normal(loc=0.2, scale=0.1, size=30),
    "model3": rng.normal(loc=0.4, scale=0.1, size=30),
    "model4": rng.normal(loc=0.5, scale=0.1, size=30),
    "model5": rng.normal(loc=0.7, scale=0.1, size=30),
    "model6": rng.normal(loc=0.7, scale=0.1, size=30),
    "model7": rng.normal(loc=0.8, scale=0.1, size=30),
    "model8": rng.normal(loc=0.9, scale=0.1, size=30),
}

samples = np.column_stack(list(models.values()))
draw_cd_diagram(samples, labels=list(models.keys()), out_file="out.svg", title="Model comparison")

Non-significant results

If the Friedman test is not significant, the function issues a warning and returns None — no diagram is produced because the data does not support ranking the models.

API

draw_cd_diagram(
    samples,                   # 2D array-like (rows=datasets, columns=models)
    labels,                    # Sequence of model names (one per column)
    title=None,                # Optional diagram title
    out_file=None,             # Optional path to write SVG file
    fig_size=None,             # Optional (width, height) tuple in pixels
    posthoc="wilcoxon-holm",   # "wilcoxon-holm" (default) or "nemenyi"
) -> Element | None

Input formats: NumPy arrays, pandas DataFrames, or any object with a .to_numpy() / .values attribute.

Release Notes

0.0.8

• Added posthoc parameter; default switched to Wilcoxon signed-rank with Holm correction to avoid Nemenyi's pool-dependency defect (Benavoli, Corani & Mangili, JMLR 2016). The original Demšar/Nemenyi behavior is available with posthoc="nemenyi".
• The CD bar is rendered only when posthoc="nemenyi" (no single critical distance exists for Wilcoxon-Holm).

0.0.7

• Optimized Nemenyi critical-value computation with a precomputed q_alpha lookup table for k=3..100 at alpha=0.05.
• Kept a SciPy studentized_range fallback for values outside the lookup range (or different alpha), preserving behavior for all valid inputs.

0.0.6

• Red clique bars now render on top of all other SVG elements (connectors, markers, axis).

0.0.5

• Text labels no longer carry an SVG stroke, removing the bold/smudged appearance.
• font-family="sans-serif" set globally for consistent cross-platform rendering.
• Empty title no longer emits a stray <text/> node.
• Non-maximal cliques filtered out — only strictly maximal non-significance groups are drawn, eliminating near-duplicate bars.
• Classifier connectors use a single <polyline> with stroke-linejoin="miter", removing the corner notch produced by two separate lines.
• Axis markers are taller than the axis stroke so they are visible.
• Clique end-caps enlarged for legibility.
• CD-bar tick length unified with ruler tick length.
• Label-width heuristic bumped to better fit sans-serif glyphs.

0.0.4

• Non-significant groups packed into the minimum number of rows (greedy interval scheduling).
• Fixed vertical layout: title → CD bar → ruler → groups → labels (groups now drawn below the axis).

0.0.2

• Replaced hardcoded Nemenyi critical-value lookup table with SciPy's studentized_range computation.
• Updated CD diagram drawing to follow continuous rank-axis placement and first-anchor non-significant grouping.
• Improved readability for larger numbers of algorithms with adaptive multi-row label layout.

License

MIT