Grouped Partial Dependence Plots for Adverse Action Reason Code Generation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Financial services companies face challenges in generating adverse action reason codes (AARCs) due to the complexity of machine learning (ML) models, which makes it difficult to analyze and rank the inputs contributing to credit decision outcomes, thereby hindering compliance with regulations like the Equal Credit Opportunity Act (ECOA).
Innovation Solution
The method involves utilizing partial dependence plots (PDPs) to interpret ML models by grouping input variables based on correlation using a clustering algorithm, generating PDP tables for each group, and creating a ranking vector that indicates the contribution of each group to the model's output. This allows for the identification of adverse action reason codes associated with the most significant contributions to credit denial decisions.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If machine learning models are used to improve credit decision-making speed and remove human bias, then productivity and fairness are improved, but the complexity of analyzing model outputs and generating adverse action reason codes increases
Solution Approach 1:
The patent introduces partial dependence plots as an intermediary visualization tool between the complex ML model and the adverse action reason code generation process. These plots serve as a mediator that translates complex model predictions into interpretable relationships between input variables and output, enabling automated generation of AARCs without requiring manual analysis of the model's internal logic.
Solution Approach 2:
The patent utilizes visualizations (analogous to color changes in TRIZ) by generating partial dependence plots that visually represent the relationship between input variables and model outputs. These visual representations transform abstract numerical predictions into intuitive graphical patterns, making the model's decision logic transparent and enabling automated AARC generation based on visual interpretation of variable importance and interaction patterns.
2Ease of operation
If traditional statistical techniques are used to generate adverse action reason codes, then ease of analysis and compliance with ECOA is improved, but the ability to handle complex machine learning models deteriorates
Solution Approach 1:
The patent creates a universal framework using partial dependence plots that can handle multiple types of machine learning models (classification, regression, ensemble methods) while maintaining ease of analysis for adverse action reason code generation. This universal approach allows the same visualization and analysis methodology to work across different complex models, providing both ease of operation and broad model compatibility.
3Productivity
If machine learning models are used to automate credit decisions, then efficiency is improved, but the ability to explain and rank contributing factors to credit denial deteriorates
Solution Approach 1:
Partial dependence plots serve as an intermediary that preserves and reveals factor contribution information from complex ML models. By visualizing the relationship between input variables and model outputs, these plots maintain the explanatory information needed for ECOA compliance while enabling automated decision-making, thus preventing loss of interpretability information.
Solution Approach 2:
The patent replaces manual analysis mechanisms with automated visualization-based analysis. Instead of requiring manual inspection of model outputs, the system automatically generates partial dependence plots that can be programmatically interpreted to identify and rank contributing factors, substituting mechanical manual analysis with automated computational visualization.
Data Source
AI summary
A framework for interpreting machine learning models is proposed that utilizes interpretability methods to determine the contribution of groups of input variables to the output of the model. Input variables are grouped based on correlation with other input variables. The groups are identified by processing a training data set with a clustering algorithm. Once the groups of input variables are defined, partial dependent plot (PDP) tables for each group are calculated and stored in a memory, which are used for calculating scores related to each group of input variables for a given instance of the input vector processed by the model. Furthermore, Shapley Additive Explanations (SHAP) values for each group can be calculated by summing the SHAP values of the input variables for a given instance of an input vector per group. These scores can then be sorted, ranked for each interpretability method, and then combined into one hybrid ranking.