Visualization

quack.visualization provides three colorblind-safe, fully customizable Matplotlib plots to inspect quantification experiments, inspired by QuaPy's plotting module:

prevalence_plot — the diagonal plot: predicted vs. true prevalence for a target class (binary).
bias_plot — box plots of signed error (estimated - true), global or binned by true prevalence.
class_distribution_plot — bar plot of class counts/proportions in a label array.

Every function returns a matplotlib.figure.Figure — nothing is shown or saved automatically, so you decide when/how to display or persist it:

fig = prevalence_plot(...)
fig.savefig("prevalence.png", dpi=300)   # or .pdf, .svg, .jpg, ...

Note

These plots depend on matplotlib, an optional dependency. Install it with: bash pip install quack[viz]

Prevalence Plot (diagonal plot)

Plots the predicted prevalence of a target class (y-axis) against its true prevalence (x-axis). Repeated experiments for the same method are pooled and binned, showing the mean trend with an optional +/- 1 std band — analogous to QuaPy's binary_diagonal.

from sklearn.linear_model import LogisticRegression
from sklearn.svm import SVC
import numpy as np

from quack.datasets import load_uci
from quack.bag_generator import PriorShiftBagGenerator  # generate test bags across the prior spectrum
from quack.quantifiers import CC, ACC, PACC
from quack.visualization import prevalence_plot

X, y = load_uci("bc-cont")
# ... split into train/test, fit CC/ACC/PACC, and generate bags with varying priors ...

method_names, true_prevs, estim_prevs = [], [], []
for name, quantifier in [("CC", CC(LogisticRegression())), ("ACC", ACC(LogisticRegression()))]:
    quantifier.fit(X_train, y_train)
    true_bag_prevs, estim_bag_prevs = [], []
    for X_bag, y_bag in bags:  # e.g. produced by PriorShiftBagGenerator
        true_bag_prevs.append(np.mean(y_bag == 1))
        estim_bag_prevs.append(quantifier.predict(X_bag)[1])
    method_names.append(name)
    true_prevs.append(np.array(true_bag_prevs))
    estim_prevs.append(np.array(estim_bag_prevs))

fig = prevalence_plot(
    method_names, true_prevs, estim_prevs,
    class_name="malignant",
    train_prevalence=np.mean(y_train == 1),
    n_bins=21,
)
fig.savefig("bc_cont_diagonal.png", dpi=300)

A method name can repeat across several experiments (e.g. one array per fold or dataset); matching entries are merged before binning, exactly like QuaPy's convention of allowing repeated method_names.

Key parameters

Parameter	Purpose
`class_name`	Label used on axes/legend for the target class.
`train_prevalence`	Marks one or more training priors on the diagonal with a star marker.
`n_bins`	Number of true-prevalence bins used to aggregate repeated experiments.
`show_std`	Toggle the +/- 1 std shaded band.
`colors`, `markers`	Override the default colorblind-safe palette / marker cycle.
`fig_size`, `font_size`, `legend_font_size`	Layout and typography controls.
`ax`	Draw into an existing `Axes` to compose custom multi-panel figures.

Bias Plot

Shows, via box plots, the signed error estimated - true prevalence. A box centered at 0 indicates an unbiased quantifier; positive boxes indicate systematic overestimation, negative boxes underestimation.

from quack.visualization import bias_plot

# Global bias per method (QuaPy's binary_bias_global)
fig = bias_plot(method_names, true_prevs, estim_prevs, class_name="malignant")
fig.savefig("bc_cont_bias_global.png")

# Bias broken down by true-prevalence range (QuaPy's binary_bias_bins)
fig = bias_plot(
    method_names, true_prevs, estim_prevs,
    class_name="malignant", n_bins=3,
)
fig.savefig("bc_cont_bias_bins.png")

This is particularly useful to spot patterns invisible in the global view — e.g. a method trained at a balanced prior can still show positive bias for low-prevalence bags and negative bias for high-prevalence ones.

Prevalence Coverage Plot

Shows how well a set of generated bags covers the [0, 1] prevalence range for a target class — a histogram of bag counts per prevalence bin plus a rug plot marking each individual bag, so gaps in coverage are immediately visible before running a full benchmark.

from quack.bag_generator import PriorShiftBagGenerator, CovariateShiftBagGenerator
from quack.visualization import prevalence_coverage_plot

prior_gen = PriorShiftBagGenerator(n_bags=200, bag_size=100, random_state=0)
cov_gen = CovariateShiftBagGenerator(n_bags=200, bag_size=100, random_state=0)

prior_gen.to_list(X_test, y_test)
cov_gen.to_list(X_test, y_test)

fig = prevalence_coverage_plot(
    [prior_gen.sampled_prevalences_[:, 1], cov_gen.sampled_prevalences_[:, 1]],
    labels=["Prior Shift", "Covariate Shift"],
    class_name="malignant",
    train_prevalence=np.mean(y_train == 1),
)
fig.savefig("coverage.png", dpi=300)

PriorShiftBagGenerator with sampling_strategy="uniform" should show near-full coverage (close to 100%) across all bins, while CovariateShiftBagGenerator — since prevalence shift is only a side effect of the covariate shift, not directly controlled — often reveals gaps or an uneven concentration of bags. Comparing the two side by side helps decide which protocol actually exercises the prevalence range you care about for your evaluation.

Class Distribution Plot

A quick sanity check for the label distribution of a training set or test bag:

from quack.visualization import class_distribution_plot

fig = class_distribution_plot(y_train, normalize=True, title="Training set class distribution")
fig.savefig("train_distribution.png")

Set horizontal=True for datasets with many classes or long class-name labels.

Combining plots in a single figure

Every function accepts an ax argument, so you can compose a multi-panel report using standard Matplotlib subplots:

import matplotlib.pyplot as plt
from quack.visualization import prevalence_plot, bias_plot, class_distribution_plot

fig, axes = plt.subplots(1, 3, figsize=(18, 5))

prevalence_plot(method_names, true_prevs, estim_prevs, ax=axes[0], title="Diagonal Plot")
bias_plot(method_names, true_prevs, estim_prevs, ax=axes[1], title="Bias")
class_distribution_plot(y_train, ax=axes[2], title="Train Distribution")

fig.tight_layout()
fig.savefig("report.pdf")

Color palette

All plots share a colorblind-safe palette (Okabe & Ito, 2008) via quack.visualization.get_color_palette, which automatically extends to any number of methods/classes by sampling a perceptually-uniform colormap once the base palette is exhausted — no two series ever share the same color, regardless of how many quantifiers are being compared.

from quack.visualization import get_color_palette

colors = get_color_palette(12)  # 8 base colorblind-safe colors + 4 extended
fig = prevalence_plot(method_names, true_prevs, estim_prevs, colors=colors)