Real-Time Fraud Detection using SageMaker and LightGBM

In this article, we will delve into how to develop and deploy a real-time fraud detection system using PyTorch, SageMaker, and an off-the-shelf open-source model, LightGBM. Amazon SageMaker provides every developer and data scientist with the ability to build, train, and deploy machine learning models quickly. Coupled with PyTorch and LightGBM, we can deliver robust real-time fraud detection solutions.

Step 1: Set up Your Environment

Firstly, let's install all the necessary libraries. Please note that AWS CLI needs to be configured before running these commands.

pip install boto3 sagemaker pytorch torch torchvision lightgbm

Step 2: Load and Preprocess the Data

In this step, we will load and preprocess the data. For the purpose of this article, we'll use a dataset from Kaggle's Credit Card Fraud Detection.

import pandas as pd
from sklearn.model_selection import train_test_split

data = pd.read_csv('creditcard.csv')

# Normalizing and splitting the data
y = data['Class']
X = data.drop('Class', axis=1)
X_train, X_test, y_train, y_test = train_test_split(X, y, test_size=0.2, random_state=42)

# Save preprocessed data to be used in SageMaker
pd.concat([y_train, X_train], axis=1).to_csv('train.csv', index=False, header=False)
pd.concat([y_test, X_test], axis=1).to_csv('test.csv', index=False, header=False)

Step 3: Upload the Data to S3

Next, we will upload the data to an S3 bucket. This can be done by using the following code:

import sagemaker
import boto3

s3 = boto3.resource('s3')
bucket = sagemaker.Session().default_bucket()

# Upload the dataset to an S3 bucket
s3.meta.client.upload_file('train.csv', bucket, 'fraud-detection/train.csv')
s3.meta.client.upload_file('test.csv', bucket, 'fraud-detection/test.csv')

Step 4: Train the Model using LightGBM on SageMaker

In this step, we will train the LightGBM model on SageMaker.

from sagemaker import get_execution_role
from sagemaker.estimator import Estimator

# Get the SageMaker execution role
role = get_execution_role()

# Create a LightGBM estimator
estimator = Estimator(
    image_uri='your-docker-image-for-lightgbm', # You need to replace this with your own Docker image for LightGBM
        'objective': 'binary',
        'metric': 'binary_logloss',
        'num_leaves': 31,
        'learning_rate': 0.05

# Start the training job{'train': 's3://{}/fraud-detection/train.csv'.format(bucket)})

Step 5: Deploy the Model

After the model has been trained, we can deploy it using SageMaker's real-time hosting functionality.

predictor = estimator.deploy(

Step 6: Real-Time Predictions

Finally, with our model deployed on SageMaker, we can make real-time predictions as follows:

import numpy as np

test_data = np.array([[...]])  # Insert your own test data here
result = predictor.predict(test_data)
print('Fraud prediction:', 'Fraud' if result > 0.5 else 'No Fraud')

Step 7: Clean Up

To avoid incurring unnecessary costs, make sure to delete the endpoint after usage.


This guide offers a comprehensive walk-through for engineers seeking to deploy a real-time fraud detection system using PyTorch, LightGBM, and SageMaker. It involves data preprocessing, model training, and deployment of the model for real-time prediction. The use of AWS SageMaker provides a smooth transition from training to deployment, allowing for a streamlined machine learning pipeline.

Disclaimer: The LightGBM docker image has to be built and pushed to ECR. You can refer to the guide here for creating a docker image. This might involve setting up LightGBM, installing necessary libraries, and defining an entry point script to train the model. Also, please note that the cost of running this code depends on the AWS services used. Always remember to delete resources after use to avoid unnecessary charges.

What will you build?

Explore workbook templates or customize your own.

Start Building