Handling class imbalance and generating synthetic data: Worked Example — Data Manipulation and Preparation (NVIDIA-Certified Associate: Accelerated Data Science)

Handling Class Imbalance and Generating Synthetic Data: Worked Example In GPU-accelerated data science workflows, managing class imbalance and...

Handling Class Imbalance and Generating Synthetic Data: Worked Example

In GPU-accelerated data science workflows, managing class imbalance and generating synthetic data are critical for building robust machine learning models. This worked example demonstrates a practical approach using RAPIDS cuDF and related tools to address these challenges effectively.

Scenario

Suppose we have a dataset for fraud detection with two classes: fraudulent transactions (minority class) and legitimate transactions (majority class). The dataset is highly imbalanced, with only 2% fraudulent cases. Our goal is to prepare the data for model training by balancing the classes through synthetic data generation.

Step 1: Load and Inspect the Dataset

Using cuDF, we load the dataset and check class distribution.

Code snippet

import cudf

df = cudf.read_csv('transactions.csv') class_counts = df['is_fraud'].value_counts() print(class_counts)

Output:

0 (legitimate): 98,000
1 (fraudulent): 2,000

The imbalance ratio is approximately 49:1, which can bias model training.

Step 2: Separate Minority and Majority Classes

We split the dataset into minority and majority subsets.

Code snippet

df_minority = df[df['is_fraud'] == 1] df_majority = df[df['is_fraud'] == 0]

Step 3: Generate Synthetic Data Using SMOTE

To balance the classes, we apply SMOTE (Synthetic Minority Over-sampling Technique), which creates synthetic samples by interpolating between minority class instances. RAPIDS cuML provides a GPU-accelerated SMOTE implementation.

Code snippet

from cuml.preprocessing import SMOTE

features = df.drop(columns=['is_fraud']) target = df['is_fraud']

smote = SMOTE(sampling_strategy='auto', random_state=42) features_resampled, target_resampled = smote.fit_resample(features, target)

This step generates synthetic minority class samples to match the majority class size.

Step 4: Verify Balanced Dataset

Check the new class distribution after resampling.

Code snippet

import cudf

resampled_df = features_resampled.copy() resampled_df['is_fraud'] = target_resampled

print(resampled_df['is_fraud'].value_counts())

Expected output:

0 (legitimate): 98,000
1 (fraudulent): 98,000

Step 5: Proceed with Model Training

With a balanced dataset, models can learn more effectively without bias towards the majority class. The synthetic data preserves the minority class distribution and feature relationships.

Summary

Class imbalance was identified and quantified using cuDF.
Data was split into minority and majority classes.
SMOTE from cuML was used to generate synthetic minority samples on GPU.
The dataset was balanced, improving conditions for model training.

This example highlights how GPU-accelerated libraries in the RAPIDS ecosystem enable efficient handling of class imbalance and synthetic data generation, key skills validated in the NVIDIA-Certified Associate: Accelerated Data Science exam.

More in this topic

Ready to test your knowledge?

Put what you've learned into practice with a quick quiz and track your progress.

Test your knowledge →