Coaching Tool Guides Rejected Loan Applicants Toward Better Outcomes
A new web-based tool is set to provide people with unprecedented visibility into the machine learning models that are used to make high-stakes decisions impacting their daily lives.
Developed at Georgia Tech, GAM Coach is the first interactive tool of its kind to give people with rejected loan applications the power to personalize recourse options. This level of personalization means the recourse options generated by GAM Coach are realistically actionable for these applicants, which can help to ensure better outcomes for them in the future.
Existing machine learning (ML) models generate recourse options based on fixed assumptions about a broad spectrum of people. GAM Coach, however, lets users iteratively adjust loan application features, such as loan amount, payment terms, credit score, homeownership status, and more, based on their personal preferences.
“We can’t assume that developers can make the best decisions for everyone,” said Zijie (Jay) Wang, lead researcher and a Ph.D. student in Georgia Tech’s School of Computational Science and Engineering (CSE).
“Our goal is to give agency to the end user, so we developed GAM Coach to give people actionable recourse in scenarios like loan applications.”
GAM Coach lets users exercise this agency by developing up to five recourse plans at a time. They can customize each iteration by adjusting sliders to set acceptable ranges for loan amount, revolving balance, and similar variable features. Emojis with related text like, “☹ Very hard to change”, are used to set difficulty levels for features that might be easier or harder to change depending on the individual applicant.
The open-source tool is designed to let users “play around,” says Wang, to see how adjusting one feature can impact the model’s prediction. ’What if I raise my FICO score 10 points?’ ‘What if I reduce the loan amount?’ ‘What if I had 10% less debt?’
“Not every option is actionable for every person, but by allowing users to interact directly with their variable preferences, GAM Coach can find the minimal number of changes an individual needs to increase the likelihood of being approved for a loan,” said Wang, a recipient of the 2023 Apple Scholars in AI/ML PhD fellowship.
If the initial five plans aren’t satisfactory, users can continue to iteratively fine-tune their recourse options until they find a plan that best meets their needs.
To build a tool that can generate personalized recourse options that are realistically actionable, Wang and his collaborators first developed an innovative new linear integer algorithm and an easy-to-use interactive data visualization interface. These were paired and then put under the hood of a generalized additive model (GAM) to create GAM Coach.
GAMs are relatively common predictive ML models that are well-suited for determining optimal solutions. They’re also known for their simplicity and transparency, which is a big reason why Wang turned to the model for this work.
“Ultimately, we want to make artificial intelligence and machine learning systems more transparent and understandable for non-technical users so, we wanted GAM Coach to be a glass box rather than a black box tool,” said Wang.
“We want people to be able to understand how and why a machine learning model makes a certain decision. We tailored our algorithm to integrate into a GAM because it is highly accurate, we know how it works, and we know how exactly how it makes predictions.”
Wang and his collaborators conducted an online user study of GAM Coach as part of the project. The team examined user logs from 41 Amazon Mechanical Turk workers to determine how everyday users would interact with the tool. The workers were presented with different loan scenarios and challenges, and then asked to use the tool to find recourse options that met their needs.
Along with a few minor usability issues, the researchers found that that personalized recourse plans are preferred over generic plans. They also found that users had a deeper understanding of how a decision was made and what they could do to change the outcome in the future.
Despite the success of the tool so far, Wang says his team would need input from financial and legal experts before GAM Coach could be used in the real world. However, a demo and the code are available.
“Developers can also use our flexible Python library (`pip install gamcoach`) to generate recourse plans for GAMs,” said Wang, who is advised by School of CSE Associate Professor Polo Chau.
He is the lead author of GAM Coach: Towards Interactive and User-centered Algorithmic Recourse. The paper has been accepted and is being presented at the 2023 ACM CHI Conference on Human Factors in Computing Systems later this month in Hamburg, Germany.