cdf_threshold_shift.py

import numpy as np

from sklearn.preprocessing import QuantileTransformer

class cdf_threshold_shift:
    def __init__(self, shift=0.0):
        self.transformer = QuantileTransformer()
        self.threshold = None
        self.shift = shift

    def __call__(self, supports, ignore_shift=False):
        if ignore_shift:
            return self.transformer.transform(supports)
        # supports have shape [num_users, num_data_points] or [num_data_points]
        transformed_supports = self.transformer.transform(supports)
        # Shift only values that are below threshold
        transformed_supports[transformed_supports < self.threshold] += self.shift
        return transformed_supports

    def train(self, data_points):
        transformed_data_points = self.transformer.fit_transform(data_points)
        self.threshold = np.percentile(np.unique(transformed_data_points), 50)
        print(f"@@@ Threshold: {self.threshold}")