I am using the Kaggle's credit card fraud detection dataset (https://www.kaggle.com/mlg-ulb/creditcardfraud)
In order to create a balanced datasets I was testing RandomUnderSampler() and NearMiss(). I am running a make_pipeline() from imblearn. I get very different results when I used RobustScaler() before vs after Neamiss() method. This drastic difference with LinearSVC(). Is this something wrong here, it is expected?
import pandas as pd
from sklearn import model_selection, preprocessing
from imblearn import under_sampling
from imblearn.pipeline import make_pipeline
credit = pd.read_csv() # Please use data from above link
X_train, X_test, y_train, y_test = model_selection.train_test_split(credit.drop('Class', 1), credit.Class, test_size = 0.2, random_state = 100)
# 1)
pipe = make_pipeline(under_sampling.NearMiss(), preprocessing.RobustScaler(), LinearSVC(dual = False))
score = model_selection.cross_val_score(pipe, X_train, y_train, cv = 3)
print(score, '\n', score.mean())
"""
results are
[0.88720062 0.8111471 0.81310897]
0.8371522304707543 """
# 2)
pipe = make_pipeline(preprocessing.RobustScaler(), under_sampling.NearMiss(), LinearSVC(dual = False))
score = model_selection.cross_val_score(pipe, X_train, y_train, cv = 3)
print(score, '\n', score.mean())
"""
results-
[0.33242044 0.46160531 0.35399221]
0.38267265137357126
"""
# 3)
pipe = make_pipeline(under_sampling.RandomUnderSampler(), preprocessing.RobustScaler(), LinearSVC(dual = False))
score = model_selection.cross_val_score(pipe, X_train, y_train, cv = 3)
print(score, '\n', score.mean())
"""
results-
[0.97021686 0.96007795 0.9701638 ]
0.966819533466512
"""
# 4)
pipe = make_pipeline(preprocessing.RobustScaler(), under_sampling.RandomUnderSampler(), LinearSVC(dual = False))
score = model_selection.cross_val_score(pipe, X_train, y_train, cv = 3)
print(score, '\n', score.mean())
"""
results-
[0.97234987 0.96139464 0.95404751]
0.9625973369943192
"""
