Integrating Process Orchestration with Machine Learning for Bank Loan Applications

Editor’s note: This post was originally published here, but we’re happy to share this contribution from our community with a broader audience. If you’d like to share a contribution, find out more about how you can contribute too.

Acknowledgment: Thanks to Camunda’s Connector team for introducing the Amazon SageMaker Connector. In a previous article, I explored the concept of Prediction as a Service using Camunda 8’s Lambda Connector. At that time, the Amazon SageMaker connector had been released, but I couldn’t find it in the Connector marketplace—possibly because it was just getting started or wasn’t fully available.

In this article, we explore the extensive use of Amazon SageMaker along with other Connectors such as the SQL Connector and Slack Connector. Additionally, we will utilize the Intermediate Message Catch Event.

Story Telling

As a Bank Manager, thoroughly evaluating applicants for loan eligibility is crucial to making sound financial decisions. Leveraging Machine Learning algorithms can significantly enhance this assessment process. Classification algorithms, such as Logistic Regression, are particularly useful for predicting whether an applicant is likely to qualify for a loan.

By applying these algorithms, we can analyze various features of an applicant’s financial history, such as credit score, income level, debt-to-income ratio, and previous loan repayment behavior. Logistic Regression helps in modeling the probability of an applicant meeting the eligibility criteria based on these features.

Additionally, other classification techniques like Decision Trees, Random Forests, and Support Vector Machines can further refine predictions. These models can be trained on historical data to identify patterns, making more accurate predictions about future applicants.

Integrating these Machine Learning models into our loan assessment process provides a more objective and data-driven approach, potentially reducing the risk of defaults and improving overall loan approval efficiency.

For this use case, I utilized Amazon SageMaker to train a sophisticated machine learning model using historical data. SageMaker’s robust platform enabled me to develop and fine-tune the model effectively. Once trained, the model was deployed as a real-time endpoint, allowing for seamless integration into our loan assessment workflow.

This deployment ensures that we can leverage the model’s predictive capabilities to evaluate applicants efficiently and accurately. By harnessing SageMaker’s powerful infrastructure, we can continuously refine the model with new data, ensuring it adapts to evolving financial trends and patterns, ultimately enhancing our decision-making process for loan approvals. However, in this article we are not covering the deployment process.

Modelling and Implementation

The BPM (Business Process Management) modeling for this use case is straightforward and does not require complex modeling techniques. Once a loan application is submitted, a process instance is triggered. This instance then calls the inference endpoint created in Amazon SageMaker. The endpoint evaluates the applicant’s data and returns a classification indicating their loan eligibility.

Configuration for Camunda Amazon SageMaker Connector:

• IAM User: Use the IAM user that has access to the model and update the Access Key and Secret Key. Also, update the region appropriately.

• Endpoint Name: Update the endpoint name, and payload that’s needed by the model to make predictions (I have stored the applicant’s data as a process variable, which can be accessed via FEEL expressions).

If you have already deployed your model for inference in SageMaker, you can find it in AWS console under SageMaker Service.

• Update target model

Target Model Name can be obtained from AWS console under Inference section of Sagemaker.

Classification and Inference

The model classifies each applicant into one of four categories:

0: Rejection due to Low Credit Score
1: Rejection due to High Debt
2: Rejection due to Missing Information
3: Approval

The inference response, which could be 0, 1, 2, or 3, is then processed by the DMN (Decision Model and Notation). The DMN interprets the numeric response and outputs the corresponding classification type.

The DMN is represented as below:

Once output from DMN is obtained a conditional gateway would decide the route to Slack notification.

Based on the output from the DMN, we notify applicants of their application status. In this example, notifications are sent via Slack.

Use of Camunda for This Case

Camunda’s platform provides an efficient way to orchestrate and automate the loan application process without requiring extensive coding. By leveraging Camunda’s Connectors and BPMN (Business Process Model and Notation) features, we can easily integrate machine learning predictions and decision-making workflows.

Key Benefits:

• No Extensive Coding Required: With Camunda, most of the process orchestration can be configured using graphical BPMN models and built-in Connectors, minimizing the need for custom code.
• Integration with Amazon SageMaker: Using Camunda’s SageMaker connector, we can seamlessly call machine learning inference endpoints to classify loan applicants based on their data.
• Decision Automation: The DMN (Decision Model and Notation) can be used to automatically interpret the machine learning model’s output and determine the applicant’s eligibility status.
• Notification Integration: Camunda’s connectors, like the Slack/Whatsapp connector, allow for easy setup of notifications to communicate the decision to applicants.

Process Overview:

1. Application Submission: A process instance is triggered upon loan application submission.
2. Inference Call: The process instance makes a call to the Amazon SageMaker inference endpoint.
3. DMN Evaluation: The response from SageMaker is processed by the DMN to determine the classification type.
4. Notification: Based on the DMN output, a conditional gateway routes the process to send a notification via Slack.

This streamlined process leverages Camunda’s powerful workflow automation capabilities, enabling quick and accurate loan application processing with minimal manual intervention.

Read next

Looking for more about how process orchestration can help in scenarios just like this? Be sure to check out our blog post, Effective Loan Automation with Process Orchestration.