Source code for empulse.metrics.lift

import numpy as np

from .._types import FloatArrayLike, FloatNDArray
from ._validation import _check_binary, _check_fraction, _check_shape, _check_variance, _check_y_pred, _check_y_true


def _validate_input(
    y_true: FloatArrayLike, y_score: FloatArrayLike, fraction: float
) -> tuple[FloatNDArray, FloatNDArray]:
    y_true: FloatNDArray = _check_y_true(y_true)
    y_score: FloatNDArray = _check_y_pred(y_score)
    _check_shape(y_true, y_score)
    _check_binary(y_true)
    _check_variance(y_true)
    _check_shape(y_true, y_score)
    _check_fraction(fraction, 'fraction')

    return y_true, y_score




def lift_score(
    y_true: FloatArrayLike, y_score: FloatArrayLike, *, fraction: float = 0.1, check_input: bool = True
) -> float:
    """
    Compute the lift score for the top fraction of predictions.

    Parameters
    ----------
    y_true : 1D array-like, shape=(n_samples,)
        Binary target values ('positive': 1, 'negative': 0).

    y_score : 1D array-like, shape=(n_samples,)
        Target scores, can either be probability estimates or non-thresholded decision values.

    fraction : float, optional, default: 0.1
        Fraction of data to consider. Must be between 0 and 1.

    check_input : bool, default=True
        Perform input validation.
        Turning off improves performance, useful when using this metric as a loss function.

    Returns
    -------
    lift_score : float
        Lift score for the top fraction of the data.
    """
    if check_input:
        y_true, y_score = _validate_input(y_true, y_score, fraction)
    else:
        y_true = np.asarray(y_true)
        y_score = np.asarray(y_score)

    # Sort the predictions in descending order
    sorted_indices = np.argsort(y_score)[::-1]
    sorted_labels = y_true[sorted_indices]
    top_fraction = int(round(len(sorted_labels) * fraction, 0))

    n_positives_top_fraction: int = int(np.sum(sorted_labels[:top_fraction]))
    prop_positives_top_fraction = n_positives_top_fraction / top_fraction

    return prop_positives_top_fraction / float(np.mean(y_true))