In Machine Learning and Data Science we often come across a term called Imbalanced Data Distribution, generally happens when observations in one of the class are much higher or lower than the other classes. As Machine Learning algorithms tend to increase accuracy by reducing the error, they do not consider the class distribution. This problem is prevalent in examples such as Fraud Detection, Anomaly Detection, Facial recognition etc.
Standard ML techniques such as Decision Tree and Logistic Regression have a bias towards the majority class, and they tend to ignore the minority class. They tend only to predict the majority class, hence, having major misclassification of the minority class in comparison with the majority class. In more technical words, if we have imbalanced data distribution in our dataset then our model becomes more prone to the case when minority class has negligible or very lesser recall.
Imbalanced Data Handling Techniques: There are mainly 2 mainly algorithms that are widely used for handling imbalanced class distribution.
1. SMOTE
2. Near Miss Algorithm
SMOTE (Synthetic Minority Oversampling Technique) – Oversampling
SMOTE (synthetic minority oversampling technique) is one of the most commonly used oversampling methods to solve the imbalance problem.
It aims to balance class distribution by randomly increasing minority class examples by replicating them.
SMOTE synthesises new minority instances between existing minority instances. It generates the virtual training records by linear interpolation for the minority class. These synthetic training records are generated by randomly selecting one or more of the k-nearest neighbors for each example in the minority class. After the oversampling process, the data is reconstructed and several classification models can be applied for the processed data.
More Deep Insights of how SMOTE Algorithm work !
• Step 1: Setting the minority class set A, for each , the k-nearest neighbors of x are obtained by calculating the Euclidean distance between x and every other sample in set A.
• Step 2: The sampling rate N is set according to the imbalanced proportion. For each , N examples (i.e x1, x2, …xn) are randomly selected from its k-nearest neighbors, and they construct the set .
• Step 3: For each example (k=1, 2, 3…N), the following formula is used to generate a new example:
in which rand(0, 1) represents the random number between 0 and 1.
NearMiss Algorithm – Undersampling
NearMiss is an under-sampling technique. It aims to balance class distribution by randomly eliminating majority class examples. When instances of two different classes are very close to each other, we remove the instances of the majority class to increase the spaces between the two classes. This helps in the classification process.
To prevent problem of information loss in most under-sampling techniques, near-neighbor methods are widely used.
The basic intuition about the working of near-neighbor methods is as follows:
• Step 1: The method first finds the distances between all instances of the majority class and the instances of the minority class. Here, majority class is to be under-sampled.
• Step 2: Then, n instances of the majority class that have the smallest distances to those in the minority class are selected.
• Step 3: If there are k instances in the minority class, the nearest method will result in k*n instances of the majority class.
For finding n closest instances in the majority class, there are several variations of applying NearMiss Algorithm :
1. NearMiss – Version 1 : It selects samples of the majority class for which average distances to the k closest instances of the minority class is smallest.
2. NearMiss – Version 2 : It selects samples of the majority class for which average distances to the k farthest instances of the minority class is smallest.
3. NearMiss – Version 3 : It works in 2 steps. Firstly, for each minority class instance, their M nearest-neighbors will be stored. Then finally, the majority class instances are selected for which the average distance to the N nearest-neighbors is the largest.
This article helps in better understanding and hands-on practice on how to choose best between different imbalanced data handling techniques.
Load libraries and data file
The dataset consists of transactions made by credit cards. This dataset has 492 fraud transactions out of 284, 807 transactions. That makes it highly unbalanced, the positive class (frauds) account for 0.172% of all transactions.
The dataset can be downloaded from here.
# import necessary modules
import pandas as pd
import matplotlib.pyplot as plt
import numpy as np
from sklearn.linear_model import LogisticRegression
from sklearn.preprocessing import StandardScaler
from sklearn.metrics import confusion_matrix, classification_report
# load the data set
data = pd.read_csv('creditcard.csv')
# print info about columns in the dataframe
print(data.info())
Output:
RangeIndex: 26808 entries, 0 to 284806
Data columns (total 31 columns):
Time 26808 non-null float64
V1 26808 non-null float64
V2 26808 non-null float64
V3 26808 non-null float64
V4 26808 non-null float64
V5 26808 non-null float64
V6 26808 non-null float64
V7 26808 non-null float64
V8 26808 non-null float64
V9 26808 non-null float64
V10 26808 non-null float64
V11 26808 non-null float64
V12 26808 non-null float64
V13 26808 non-null float64
V14 26808 non-null float64
V15 26808 non-null float64
V16 26808 non-null float64
V17 26808 non-null float64
V18 26808 non-null float64
V19 26808 non-null float64
V20 26808 non-null float64
V21 26808 non-null float64
V22 26808 non-null float64
V23 26808 non-null float64
V24 26808 non-null float64
V25 26808 non-null float64
V26 26808 non-null float64
V27 26808 non-null float64
V28 26808 non-null float64
Amount 26808 non-null float64
Class 26808 non-null int64
# normalise the amount column
data['normAmount'] = StandardScaler().fit_transform(np.array(data['Amount']).reshape(-1, 1))
# drop Time and Amount columns as they are not relevant for prediction purpose
data = data.drop(['Time', 'Amount'], axis = 1)
# as you can see there are 492 fraud transactions.
data['Class'].value_counts()
Output:
0 284315
1 492
Split the data into test and train sets
from sklearn.model_selection import train_test_split
# split into 70:30 ration
x-train, X-test, y-train, y-test = train_test_split(X, y, test-size = 0.3, random-state = 0)
# describes info about train and test set
print("Number transactions x-train dataset: ", x-train.shape)
print("Number transactions y-train dataset: ", y-train.shape)
print("Number transactions X-test dataset: ", X-test.shape)
print("Number transactions y-test dataset: ", y-test.shape)
Output:
Number transactions x-train dataset: (199364, 29)
Number transactions y-train dataset: (199364, 1)
Number transactions X-test dataset: (85443, 29)
Number transactions y-test dataset: (85443, 1)
Now train the model without handling the imbalanced class distribution
# logistic regression object
lr = LogisticRegression()
# train the model on train set
lr.fit(x-train, y-train.ravel())
predictions = lr.predict(X-test)
# print classification report
print(classification_report(y-test, predictions))
Output:
precision recall f1-score support
0 1.00 1.00 1.00 85296
1 0.88 0.62 0.73 147
accuracy 1.00 85443
macro avg 0.94 0.81 0.86 85443
weighted avg 1.00 1.00 1.00 85443
The accuracy comes out to be 100% but did you notice something strange ?
The recall of the minority class in very less. It proves that the model is more biased towards majority class. So, it proves that this is not the best model.
Now, we will apply different imbalanced data handling techniques and see their accuracy and recall results.
Using SMOTE Algorithm
You can check all the parameters from here.
print("Before OverSampling, counts of label '1': {}".format(sum(y-train == 1)))
print("Before OverSampling, counts of label '0': {} n".format(sum(y-train == 0)))
# import SMOTE module from imblearn library
# pip install imblearn (if you don't have imblearn in your system)
from imblearn.over_sampling import SMOTE
sm = SMOTE(random-state = 2)
x-train_res, y-train_res = sm.fit_sample(x-train, y-train.ravel())
print('After OverSampling, the shape of train_X: {}'.format(x-train_res.shape))
print('After OverSampling, the shape of train_y: {} n'.format(y-train_res.shape))
print("After OverSampling, counts of label '1': {}".format(sum(y-train_res == 1)))
print("After OverSampling, counts of label '0': {}".format(sum(y-train_res == 0)))
Output:
Before OverSampling, counts of label '1': [345]
Before OverSampling, counts of label '0': [199019]
After OverSampling, the shape of train_X: (398038, 29)
After OverSampling, the shape of train_y: (398038, )
After OverSampling, counts of label '1': 199019
After OverSampling, counts of label '0': 199019
Look! that SMOTE Algorithm has oversampled the minority instances and made it equal to majority class. Both categories have equal amount of records. More specifically, the minority class has been increased to the total number of majority class.
Now see the accuracy and recall results after applying SMOTE algorithm (Oversampling).
Prediction and Recall
lr1 = LogisticRegression()
lr1.fit(x-train_res, y-train_res.ravel())
predictions = lr1.predict(X-test)
# print classification report
print(classification_report(y-test, predictions))
Output:
precision recall f1-score support
0 1.00 0.98 0.99 85296
1 0.06 0.92 0.11 147
accuracy 0.98 85443
macro avg 0.53 0.95 0.55 85443
weighted avg 1.00 0.98 0.99 85443
Wow, We have reduced the accuracy to 98% as compared to previous model but the recall value of minority class has also improved to 92 %. This is a good model compared to the previous one. Recall is great.
Now, we will apply NearMiss technique to Under-sample the majority class and see its accuracy and recall results.
NearMiss Algorithm:
You can check all the parameters from here.
print("Before Undersampling, counts of label '1': {}".format(sum(y-train == 1)))
print("Before Undersampling, counts of label '0': {} n".format(sum(y-train == 0)))
# apply near miss
from imblearn.under_sampling import NearMiss
nr = NearMiss()
x-train_miss, y-train_miss = nr.fit_sample(x-train, y-train.ravel())
print('After Undersampling, the shape of train_X: {}'.format(x-train_miss.shape))
print('After Undersampling, the shape of train_y: {} n'.format(y-train_miss.shape))
print("After Undersampling, counts of label '1': {}".format(sum(y-train_miss == 1)))
print("After Undersampling, counts of label '0': {}".format(sum(y-train_miss == 0)))
Output:
Before Undersampling, counts of label '1': [345]
Before Undersampling, counts of label '0': [199019]
After Undersampling, the shape of train_X: (690, 29)
After Undersampling, the shape of train_y: (690, )
After Undersampling, counts of label '1': 345
After Undersampling, counts of label '0': 345
The NearMiss Algorithm has undersampled the majority instances and made it equal to majority class. Here, the majority class has been reduced to the total number of minority class, so that both classes will have equal number of records.
Prediction and Recall
# train the model on train set
lr2 = LogisticRegression()
lr2.fit(x-train_miss, y-train_miss.ravel())
predictions = lr2.predict(X-test)
# print classification report
print(classification_report(y-test, predictions))
Output:
precision recall f1-score support
0 1.00 0.56 0.72 85296
1 0.00 0.95 0.01 147
accuracy 0.56 85443
macro avg 0.50 0.75 0.36 85443
weighted avg 1.00 0.56 0.72 85443
This model is better than the first model because it classifies better and also the recall value of minority class is 95 %. But due to undersampling of majority class, its recall has decreased to 56 %. So in this case, SMOTE is giving me a great accuracy and recall, I’ll go ahead and use that model!