A Submodular Optimization Approach to Accountable Loan Approval

Kyungsik Lee, Hana Yoo, Sumin Shin, Wooyoung Kim, Yeonung Baek, Hyunjin Kang, Jaehyun Kim, Kee-Eung KimAAAI 2024 · DOI

In the financial sector, "underwriting"—the process of determining loan eligibility—is a critical balancing act. Lenders strive to maximize loan volume to drive revenue while simultaneously minimizing the "bad rate," or the frequency of defaults. While advanced machine learning models like XGBoost or Deep Neural Networks (DNNs) offer high predictive accuracy, they often function as "black boxes." In many jurisdictions, including South Korea, regulations and consumer rights mandates require loan decisions to be both accountable and interpretable. When a customer is denied credit, the institution must provide a clear, simple explanation.

Historically, this requirement led to human analysts manually crafting binary decision rules (e.g., "Reject if credit score < 700"). However, manual rule-making is inherently slow, struggles to scale with hundreds of variables, and often results in overly conservative strategies that leave potential revenue on the table. To address this, researchers at Hyundai Capital Services (HCS) and KAIST developed an automated system that leverages submodular optimization to generate "white-box" loan approval rules that rival the performance of complex models.

The HCS Automated Loan Approval Rule Engine

The proposed system bridges the gap between high-performance machine learning and human readability through a sophisticated three-stage pipeline.

  1. Search Space Builder: Rather than relying on a single complex model, the system uses XGBoost to generate thousands of "atomic rules"—simple logical conditions such as if Feature A > x AND Feature B < y. These rules are extracted from the branches of decision trees. To ensure these rules are intuitive and legally defensible, the authors enforce monotonicity. For instance, an increase in past delinquencies must always lead to a higher probability of rejection, never a lower one.
  2. Rule Construction Engine: Selecting the optimal subset of rules from a pool of tens of thousands is a massive combinatorial challenge. The researchers framed this as a submodular optimization problem. Submodularity effectively manages "diminishing returns": adding a second rule that identifies the same high-risk customers as the first provides less marginal value than a rule targeting a different segment. A greedy algorithm is employed to select a set of rules that maximizes the detection of "bad" customers within a specific "volume loss" budget.
  3. Human-in-the-Loop Simulation: The system includes a toolkit for risk analysts to visualize the trade-offs between risk and volume. This allows human experts to set final cut-off points based on prevailing market conditions and institutional risk appetite.
The end-to-end system architecture of the HCS Automated Loan Approval Rule Engine, integrating big data platforms with MLOps for strategy deployment.

Data and Feature Engineering

The study utilized real-world data from HCS personal loan customers as of August 2020. The initial dataset comprised 1,656 features, including internal credit scores, credit bureau data, and digital weblogs. Through a rigorous selection process focusing on both predictive power and explainability, this was refined to 358 features. Customers were labeled "bad" if they experienced a delinquency of 60 days or more within a 12-month period. To ensure robustness against changing market trends, the model was evaluated on "out-of-time" data from a subsequent period.

Experimental Results and Business Impact

The "SUBMOD" approach demonstrated that interpretability does not require sacrificing performance. It achieved an F1-score of 6.6%, nearly matching the 6.7% of "black-box" XGBoost and significantly outperforming the 5.2% human baseline. By treating rule selection as a set-covering problem, the engine identified overlaps and hidden patterns that human analysts typically miss. For example, it uncovered non-obvious indicators of creditworthiness, such as consistent insurance fee payments or recent vehicle purchases.

Performance comparison of the SUBMOD approach against the human expert baseline, showing significant gains in loan volume for the same risk level.

Since its deployment, HCS has seen a 14% annual growth in credit loan volume. Furthermore, the system’s precision—its ability to correctly identify bad loans—improved by approximately 29% compared to the previous human-expert baseline. Beyond performance gains, the automation of this process has enabled greater agility, allowing for "soft-retraining" every three months to respond to economic shifts.

Conclusion and Future Directions

While highly effective, the system faces challenges in high-risk segments (e.g., customers with credit scores below 700), where automated decisions require higher levels of stakeholder buy-in. Future work will focus on integrating "falling rule lists" and expanding the engine's application to debt collection strategies, prioritizing interventions like AI call-bots versus human agents based on recovery likelihood. This research proves that submodular optimization can effectively automate complex, regulated decision-making processes, providing a blueprint for accountable AI in finance.

Summary updated

BibTeX
@inproceedings{lee2024submodular,
  author = {Kyungsik Lee and Hana Yoo and Sumin Shin and Wooyoung Kim and Yeonung Baek and Hyunjin Kang and Jaehyun Kim and Kee-Eung Kim},
  title = {{A Submodular Optimization Approach to Accountable Loan Approval}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2024},
  doi = {10.1609/aaai.v38i21.30310},
}
Figure