Transforming Fraud Detection with Radial’s Modernized ML Workflow

This post is co-written with Qing Chen and Mark Sinclair from Radial. Radial, the largest 3PL fulfillment provider, not only excels in logistics but also offers integrated payment solutions, fraud detection, and omnichannel services tailored to mid-market and enterprise brands. With over 30 years of experience, Radial is committed to aligning its services with each brand’s unique needs.

Navigating the E-commerce Landscape with Confidence

In an era where e-commerce has become a cornerstone of retail, Radial helps brands address common challenges. These range from providing scalable, flexible fulfillment solutions for delivery consistency to ensuring secure transactions. With a commitment to fulfilling promises from "click to delivery," Radial empowers brands to navigate the ever-changing digital landscape by delivering a seamless, secure, and superior e-commerce experience.

The Necessity for Advanced Fraud Detection Models

Machine learning (ML) has proven to be a game-changer in the field of fraud detection compared to traditional methods. Unlike conventional approaches, ML models can sift through vast amounts of transactional data, absorbing historical fraud patterns and identifying anomalies that may signal potential threats in real time. One of the primary benefits of ML is its ability to continuously adapt and learn from new fraud tactics, ensuring robust and resilient detection systems that enhance accuracy and reduce false positives over time.

To modernize their fraud detection capabilities, Radial chose to migrate and optimize their ML workflow using Amazon SageMaker. By leveraging the AWS Experience-Based Acceleration (EBA) program, they aimed to boost efficiency, scalability, and maintainability through close collaboration.

Challenges Faced with On-Premises ML Models

While ML offers significant benefits in combating evolving fraud patterns, maintaining these models on premises poses distinct challenges, particularly regarding scalability and upkeep.

Scalability Issues

On-premises systems are limited by the physical hardware available, making them unable to cope with sudden spikes in transaction volumes—especially during peak shopping seasons. This often results in slower processing times and a diminished capacity to run multiple ML applications simultaneously, potentially leading to missed fraud detections. Moreover, scaling up an on-premises infrastructure can be a costly and lengthy endeavor, creating bottlenecks for data scientists who may have to wait for availability or reduce the scope of their experiments. This not only delays innovation but also compromises model performance.

Maintenance Complexities

The upkeep of on-premises infrastructure for fraud detection requires a dedicated IT team responsible for managing myriad tasks, from servers and storage to networking and backups. Regular retraining of fraud detection models is necessary due to natural performance degradation over time. On-premises systems often lack built-in automation tools to manage the full ML lifecycle, leading to increased operational complexity and a higher likelihood of errors.

Modernization Challenges in ML Cloud Migration

Organizations face several challenges when modernizing their ML workloads via cloud migration.

Skill Gaps

A considerable hurdle is the lack of expertise among developers and data scientists in microservices architecture and advanced ML tools suitable for cloud environments. This knowledge gap can lead to costly architectures and increased security vulnerabilities.

Cross-Functional Barriers

Limited communication across teams can impede modernization efforts. When different departments fail to share information, decision-making slows down, causing organizations to lag behind competitors.

Project Management Complexities

Modernization initiatives often require coordination across multiple teams with conflicting priorities. Aligning stakeholders around business outcomes and quantifying the benefits to prove value can become daunting tasks.

To tackle these challenges, AWS designed the EBA program, aimed squarely at helping organizations accelerate their cloud journey by enhancing collaboration and resolving roadblocks.

AWS Experience-Based Acceleration (EBA) Program: A Pathway to Collaboration

The EBA program consists of a three-day interactive workshop that capitalizes on the potential of SageMaker to improve business outcomes. It guides participants through a prescriptive ML lifecycle, starting from identifying business goals to model development and deployment.

For Radial, which already had an existing on-premises ML infrastructure, the EBA approach involved customizing the use of SageMaker to alleviate their current operational challenges. During the workshop, AWS ML experts teamed up with Radial’s cross-functional teams to offer tailored advice and bolster their capabilities.

Evolution from Legacy to Modern ML Workflows

Before the migration to SageMaker, Radial relied on on-premises systems, facing inefficiencies in model development and deployment.

Legacy Workflow Challenges

The typical process for developing a new fraud detection model took between two to four weeks. This involved:

1. Data cleaning and exploratory data analysis (EDA)
2. Feature engineering
3. Model prototyping and training
4. Model evaluation to finalize the fraud detection model

These steps had to be carried out under hardware constraints, limiting concurrent experiment runs. Once finalized, the model artifacts would be passed to the software development and DevOps teams for deployment, adding another two to three weeks to the project.

The Transition to SageMaker and MLOps

Upon migrating to SageMaker and adopting a Machine Learning Operations (MLOps) approach, Radial streamlined their entire ML lifecycle:

Model Development: The data science team maintained their essential tasks—including data cleaning and model training—but they benefited from the on-demand computing resources of SageMaker, allowing for more concurrent training experiments.

Seamless Model Deployment: The approved model can directly trigger deployment to a test environment. The process is automated, removing the need for extensive communication with software teams and deploying models in minutes rather than weeks.

Integration and Testing: The software team can quickly integrate the model with existing systems, facilitating necessary tests like integration and load testing. Once validated, production deployment happens swiftly.

Key Components of the MLOps Architecture

With their newly modernized ML workflow established, Radial incorporated best practices in MLOps. They utilized:

• SageMaker: Facilitating essential tasks from training to deployment, with built-in monitoring capabilities.
• GitLab CI/CD: Automating workflows for testing and deployment, reducing manual overhead.
• Infrastructure as Code (IaC): Using Terraform and AWS CloudFormation to manage AWS resources reliably.

Radial adopted a multi-account strategy for robust security and streamlined operations. Each account isolates environments to enforce strict security boundaries and promote efficient collaboration across teams.

Data Privacy and Compliance

Radial prioritized security and compliance, particularly as their fraud detection ML APIs process sensitive information like transaction details. To ensure they met stringent regulations like CPPA and PCI, they leveraged AWS technologies for secure data management:

• AWS Direct Connect: Establishing a dedicated high-speed connection to transfer sensitive data securely.
• Amazon VPC: Isolating environments in private subnets to enhance security.
• AWS KMS: Enforcing encryption for data at rest in Amazon S3, also providing robust data retention policies.

Benefits of the New ML Workflow on AWS

The newly implemented ML workflow brought a multitude of advantages:

• Dynamic Scalability: AWS allows Radial to react quickly to transaction spikes.
• Faster Provisioning: Model deployment cycles reduced from weeks to mere minutes.
• Consistent Model Deployment: Streamlined operations promote reliable transitions from development to production.
• Built-in Monitoring: Continuous tracking of model performance facilitates timely adjustments.

Key Takeaways for Organizational Modernization

Radial’s migration journey provides several lessons for organizations considering a similar transformation:

• Collaborate with AWS: Engage for customized solutions that fit specific use cases.
• Iterative Customization: Maintain ongoing communication with AWS Support to adapt solutions over time.
• Account Isolation Strategies: Enhancing collaboration while ensuring security through environment separation is vital.
• Fine-Tune Scaling Metrics: Regular load testing will inform any necessary adjustments to scaling configurations.

By adopting such strategies, organizations can not only respond effectively to emerging challenges but also harness the full potential of their ML capabilities across numerous applications.