This guide walks you through a instalation, training and testing a classifier-based quantifier, and evaluating the predictions using a quack's built-in metrics.

Note: this page may be adjusted in the feature as we are in a alpha version of the package, if you use it in any of you projects be careful to clone, install and use a 'stable' version (main branch). If you want to keep up with the updates you can use the 'develop' branch or some specific branch for other features.

Instalation

For now, to install quack you need to rely on the git pip feature. Just need to run the following command:

pip install git+https://github.com/jose-gilberto/quack@main

Quickstart Workflow

Considering that you already have a dataset loaded and splitted into training and testing, the following script will create a quantifier based on logistic regression to estimate the prevalence of your set.

Any data compatible with Scikit-learn or Numpy can be used, whether np.ndarray or sparse_matrix objects.

To evaluate the method, we used the Absolute Error (AE) between the predicted and real prevalences.

from sklearn.linear_model import LogisticRegression
import numpy as np
# Import from quack
from quack.quantifiers import CC
from quack.metrics import ae

# 1. Load the dataset and count the true prevalence
X_train, y_train = ...
X_test, y_test = ...

labels, counts = np.unique(y_test, return_counts=True)
true_prev = counts / y_test.shape[0]
true_prev

# 2. Initialize and fit a Classify & Count (CC) quantifier
quantifier = CC(classifier=LogisticRegression())
quantifier.fit(X, y)

predicted_prev = quantifier.predict(X_test)
# Measure performance using Absolute Error
error = ae(true_prev, predicted_prev)

print(f"True: {true_prev[1]:.2f} | Pred: {predicted_prev[1]:.2f} | AE: {error:.4f}")

Quantification is very useful in scenarios where prior probability shifts occur.