Scored Observation Explanations Using Partial Dependence Constraints
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Complex nonlinear scoring models are difficult to explain, leading to a shortfall in providing concise explanations for scored observations, which is crucial for stakeholder understanding and compliance with regulations.
Innovation Solution
Implementing systems and methods that constrain scoring models using partial dependence functions (PDFs) to ensure low-dimensional explanations, such as one- or two-dimensional PDFs, by pre-computing and storing these in look-up tables, and combining them with scored observations in real-time to generate fast explanations.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If complex nonlinear scoring models are used to improve rank-ordering performance, then the model's ability to accurately rank observations is improved, but the explainability of scores deteriorates
Solution Approach 1:
The patent segments the complex nonlinear model into multiple linear models, each responsible for a specific score range. By dividing the scoring domain into segments and applying simple linear models to each segment, the system maintains high rank-ordering performance while ensuring explainability within each segment. The segmentation allows the complex problem to be broken down into manageable, explainable parts.
Solution Approach 2:
The patent introduces a new dimension by adding a 'segment identifier' to the model output. Instead of providing a single complex score, the model now provides both a score and an identifier indicating which linear segment applies. This additional dimension allows stakeholders to understand which simple linear relationship governs their specific case, thereby improving explainability without sacrificing overall ranking accuracy.
2Reliability
If complex nonlinear models are used to achieve better scoring accuracy, then the primary rank-ordering objective is improved, but the ability to provide concise explanations deteriorates
Solution Approach 1:
The patent segments the scoring function into multiple linear segments, where each segment has its own simple interpretable parameters. This segmentation allows the model to capture complex nonlinear relationships across the full score range while maintaining simple, explainable relationships within each segment. The explanation clarity is preserved because each segment can be explained using straightforward linear relationships.
Solution Approach 2:
The patent implements a dynamic model selection mechanism that automatically selects which linear segment applies to each observation based on its characteristics. This dynamic approach allows the system to adapt to different regions of the input space, providing both high accuracy (by selecting the appropriate segment) and clear explanations (by using the simple linear relationship of the selected segment).
Data Source
AI summary
Systems and methods for generating concise explanations of scored observations that strike good, and computationally efficient, trade-offs between rank-ordering performance and explainability of scored observations are disclosed. The systems and methods described herein for explaining scored observations are based on a framework of partial dependence functions (PDFs), multi-layered neural networks (MNNs), and Latent Explanations Neural Network Scoring (LENNS).