How to use sampling_strategy in imbalanced-learn

This example shows the different usage of the parameter sampling_strategy for the different family of samplers (i.e. over-sampling, under-sampling. or cleaning methods).

# Authors: Guillaume Lemaitre <g.lemaitre58@gmail.com>
# License: MIT

print(__doc__)
import seaborn as sns

sns.set_context("poster")

Create an imbalanced dataset

First, we will create an imbalanced data set from a the iris data set.

from sklearn.datasets import load_iris

from imblearn.datasets import make_imbalance

iris = load_iris(as_frame=True)

sampling_strategy = {0: 10, 1: 20, 2: 47}
X, y = make_imbalance(iris.data, iris.target, sampling_strategy=sampling_strategy)

import matplotlib.pyplot as plt

fig, axs = plt.subplots(ncols=2, figsize=(10, 5))
autopct = "%.2f"
iris.target.value_counts().plot.pie(autopct=autopct, ax=axs[0])
axs[0].set_title("Original")
y.value_counts().plot.pie(autopct=autopct, ax=axs[1])
axs[1].set_title("Imbalanced")
fig.tight_layout()

Using sampling_strategy in resampling algorithms

sampling_strategy as a float

sampling_strategy can be given a float. For under-sampling methods, it corresponds to the ratio \(\alpha_{us}\) defined by \(N_{rM} = \alpha_{us} \times N_{m}\) where \(N_{rM}\) and \(N_{m}\) are the number of samples in the majority class after resampling and the number of samples in the minority class, respectively.

# select only 2 classes since the ratio make sense in this case
binary_mask = y.isin([0, 1])
binary_y = y[binary_mask]
binary_X = X[binary_mask]

from imblearn.under_sampling import RandomUnderSampler

sampling_strategy = 0.8
rus = RandomUnderSampler(sampling_strategy=sampling_strategy)
X_res, y_res = rus.fit_resample(binary_X, binary_y)
ax = y_res.value_counts().plot.pie(autopct=autopct)
_ = ax.set_title("Under-sampling")

For over-sampling methods, it correspond to the ratio \(\alpha_{os}\) defined by \(N_{rm} = \alpha_{os} \times N_{M}\) where \(N_{rm}\) and \(N_{M}\) are the number of samples in the minority class after resampling and the number of samples in the majority class, respectively.

from imblearn.over_sampling import RandomOverSampler

ros = RandomOverSampler(sampling_strategy=sampling_strategy)
X_res, y_res = ros.fit_resample(binary_X, binary_y)
ax = y_res.value_counts().plot.pie(autopct=autopct)
_ = ax.set_title("Over-sampling")

sampling_strategy as a str

sampling_strategy can be given as a string which specify the class targeted by the resampling. With under- and over-sampling, the number of samples will be equalized.

Note that we are using multiple classes from now on.

sampling_strategy = "not minority"

fig, axs = plt.subplots(ncols=2, figsize=(10, 5))
rus = RandomUnderSampler(sampling_strategy=sampling_strategy)
X_res, y_res = rus.fit_resample(X, y)
y_res.value_counts().plot.pie(autopct=autopct, ax=axs[0])
axs[0].set_title("Under-sampling")

sampling_strategy = "not majority"
ros = RandomOverSampler(sampling_strategy=sampling_strategy)
X_res, y_res = ros.fit_resample(X, y)
y_res.value_counts().plot.pie(autopct=autopct, ax=axs[1])
_ = axs[1].set_title("Over-sampling")

With cleaning method, the number of samples in each class will not be equalized even if targeted.

from imblearn.under_sampling import TomekLinks

sampling_strategy = "not minority"
tl = TomekLinks(sampling_strategy=sampling_strategy)
X_res, y_res = tl.fit_resample(X, y)
ax = y_res.value_counts().plot.pie(autopct=autopct)
_ = ax.set_title("Cleaning")

sampling_strategy as a dict

When sampling_strategy is a dict, the keys correspond to the targeted classes. The values correspond to the desired number of samples for each targeted class. This is working for both under- and over-sampling algorithms but not for the cleaning algorithms. Use a list instead.

fig, axs = plt.subplots(ncols=2, figsize=(10, 5))

sampling_strategy = {0: 10, 1: 15, 2: 20}
rus = RandomUnderSampler(sampling_strategy=sampling_strategy)
X_res, y_res = rus.fit_resample(X, y)
y_res.value_counts().plot.pie(autopct=autopct, ax=axs[0])
axs[0].set_title("Under-sampling")

sampling_strategy = {0: 25, 1: 35, 2: 47}
ros = RandomOverSampler(sampling_strategy=sampling_strategy)
X_res, y_res = ros.fit_resample(X, y)
y_res.value_counts().plot.pie(autopct=autopct, ax=axs[1])
_ = axs[1].set_title("Under-sampling")

sampling_strategy as a list

When sampling_strategy is a list, the list contains the targeted classes. It is used only for cleaning methods and raise an error otherwise.

sampling_strategy = [0, 1, 2]
tl = TomekLinks(sampling_strategy=sampling_strategy)
X_res, y_res = tl.fit_resample(X, y)
ax = y_res.value_counts().plot.pie(autopct=autopct)
_ = ax.set_title("Cleaning")

sampling_strategy as a callable

When callable, function taking y and returns a dict. The keys correspond to the targeted classes. The values correspond to the desired number of samples for each class.

def ratio_multiplier(y):
    from collections import Counter

    multiplier = {1: 0.7, 2: 0.95}
    target_stats = Counter(y)
    for key, value in target_stats.items():
        if key in multiplier:
            target_stats[key] = int(value * multiplier[key])
    return target_stats


X_res, y_res = RandomUnderSampler(sampling_strategy=ratio_multiplier).fit_resample(X, y)
ax = y_res.value_counts().plot.pie(autopct=autopct)
ax.set_title("Under-sampling")
plt.show()

Estimated memory usage: 200 MB