"""An undersampling transformer to downsample the majority classes in the dataset."""
import numpy as np
import pandas as pd
from evalml.pipelines.components.transformers.samplers.base_sampler import BaseSampler
from evalml.utils.woodwork_utils import infer_feature_types
[docs]class Undersampler(BaseSampler):
"""Initializes an undersampling transformer to downsample the majority classes in the dataset.
This component is only run during training and not during predict.
Args:
sampling_ratio (float): The smallest minority:majority ratio that is accepted as 'balanced'. For instance, a 1:4 ratio would be
represented as 0.25, while a 1:1 ratio is 1.0. Must be between 0 and 1, inclusive. Defaults to 0.25.
sampling_ratio_dict (dict): A dictionary specifying the desired balanced ratio for each target value. For instance, in a binary case where class 1 is the minority, we could specify:
`sampling_ratio_dict={0: 0.5, 1: 1}`, which means we would undersample class 0 to have twice the number of samples as class 1 (minority:majority ratio = 0.5), and don't sample class 1.
Overrides sampling_ratio if provided. Defaults to None.
min_samples (int): The minimum number of samples that we must have for any class, pre or post sampling. If a class must be downsampled, it will not be downsampled past this value.
To determine severe imbalance, the minority class must occur less often than this and must have a class ratio below min_percentage.
Must be greater than 0. Defaults to 100.
min_percentage (float): The minimum percentage of the minimum class to total dataset that we tolerate, as long as it is above min_samples.
If min_percentage and min_samples are not met, treat this as severely imbalanced, and we will not resample the data.
Must be between 0 and 0.5, inclusive. Defaults to 0.1.
random_seed (int): The seed to use for random sampling. Defaults to 0.
Raises:
ValueError: If sampling_ratio is not in the range (0, 1].
ValueError: If min_sample is not greater than 0.
ValueError: If min_percentage is not between 0 and 0.5, inclusive.
"""
name = "Undersampler"
hyperparameter_ranges = {}
"""{}"""
def __init__(
self,
sampling_ratio=0.25,
sampling_ratio_dict=None,
min_samples=100,
min_percentage=0.1,
random_seed=0,
**kwargs,
):
if sampling_ratio <= 0 or sampling_ratio > 1:
raise ValueError(
f"sampling_ratio must be within (0, 1], but received {sampling_ratio}",
)
if min_samples <= 0:
raise ValueError(
f"min_sample must be greater than 0, but received {min_samples}",
)
if min_percentage <= 0 or min_percentage > 0.5:
raise ValueError(
f"min_percentage must be between 0 and 0.5, inclusive, but received {min_percentage}",
)
parameters = {
"sampling_ratio": sampling_ratio,
"min_samples": min_samples,
"min_percentage": min_percentage,
"sampling_ratio_dict": sampling_ratio_dict,
}
self.sampling_ratio = sampling_ratio
self.min_samples = min_samples
self.min_percentage = min_percentage
self.random_seed = random_seed
self.sampling_ratio_dict = sampling_ratio_dict or {}
parameters.update(kwargs)
super().__init__(
parameters=parameters,
component_obj=None,
random_seed=random_seed,
)
def _initialize_sampler(self, X, y):
"""Helper function to initialize the undersampler component object.
Args:
X (pd.DataFrame): Ignored.
y (pd.Series): The target data.
"""
pass
def _find_ideal_samples(self, y):
"""Returns dictionary of examples to drop for each class if we need to resample.
Arguments:
y (pd.Series): Target data passed in.
Returns:
dict: Dictionary with undersample target class as key, and number of samples to remove as the value.
If we don't need to resample, returns empty dictionary.
"""
counts = y.value_counts()
normalized_counts = y.value_counts(normalize=True)
minority_class_count = min(normalized_counts)
class_ratios = minority_class_count / normalized_counts
# if no class ratios are larger than what we consider balanced, then the target is balanced
if all(class_ratios >= self.sampling_ratio):
return {}
# if any classes have less than min_samples counts and are less than min_percentage of the total data,
# then it's severely imbalanced
if any(counts < self.min_samples) and any(
normalized_counts < self.min_percentage,
):
return {}
# otherwise, we are imbalanced enough to perform on this
undersample_classes = counts[class_ratios <= self.sampling_ratio].index.values
# find goal size, round it down if it's a float
minority_class = min(counts.values)
goal_value = max(
int((minority_class / self.sampling_ratio) // 1),
self.min_samples,
)
# we don't want to drop less than 0 rows
drop_values = {k: max(0, counts[k] - goal_value) for k in undersample_classes}
return {k: v for k, v in drop_values.items() if v > 0}
def _sampling_dict_to_remove_dict(self, y):
"""Turns the sampling dict input into a dict of samples to remove for each target, similar to the return of _find_ideal_samples.
Arguments:
y (pd.Series): Training data targets.
Returns:
(dict): dictionary with undersample target class as key, and number of samples to remove as the value.
If we don't need to resample, returns empty dictionary.
"""
y_dict = y.value_counts().to_dict()
new_dic = {}
for k, v in self.sampling_ratio_dict.items():
new_dic[k] = max(y_dict[k] - v, 0)
return new_dic
[docs] def fit_resample(self, X, y):
"""Resampling technique for this sampler.
Arguments:
X (pd.DataFrame): Training data to fit and resample.
y (pd.Series): Training data targets to fit and resample.
Returns:
list: Indices to keep for training data.
"""
if self.parameters["sampling_ratio_dict"]:
self.sampling_ratio_dict = self._convert_dictionary(
self.parameters["sampling_ratio_dict"],
y,
)
y = infer_feature_types(y)
random_state = np.random.RandomState(self.random_seed)
if len(self.sampling_ratio_dict):
result = self._sampling_dict_to_remove_dict(y)
else:
result = self._find_ideal_samples(y)
indices_to_drop = []
if len(result):
# iterate through the classes we need to undersample and remove the number of samples we need to remove
for key, value in result.items():
indices = y.index[y == key].values
indices_to_remove = random_state.choice(indices, value, replace=False)
indices_to_drop.extend(indices_to_remove)
# indices of the y series
original_indices = list(set(y.index.values).difference(set(indices_to_drop)))
return original_indices
