SHAP Explanation Repair for Non-Linear Model Accuracy
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Non-linear machine learning models, such as multi-layer perceptrons, often struggle to generate precise explanations for their decisions, which is crucial for regulatory compliance and improving model accuracy, especially in applications like credit decisioning.
Innovation Solution
A method using the SHapley Additive exPlanations (SHAP) technique to compute errors in attribute explanations, identify adjustments to minimize confidence intervals, and apply repairs to enhance explanation quality and accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a non-linear machine learning model is used to improve predictive accuracy, then model performance is improved, but the ability to generate precise explanations deteriorates
Solution Approach 1:
The patent introduces an intermediary explanation generation system that mediates between the complex non-linear model and the user. This system uses SHAP (SHapley Additive exPlanations) values to decompose the complex model's decision-making into interpretable attribute contributions, providing precise explanations without requiring changes to the underlying complex model architecture.
Solution Approach 2:
The patent replaces direct interpretation of complex non-linear model mechanics with an alternative explanation mechanism. Instead of trying to understand the intricate internal workings of the non-linear model, the system uses game-theoretic SHAP values to compute fair attribution of predictions to input features, substituting mechanical interpretation with a mathematical fairness-based approach.
2Reliability
If explanation computation is performed for non-linear models, then regulatory compliance is improved, but computational intractability increases
Solution Approach 1:
The patent performs preliminary computation of SHAP values and error estimates before final explanation generation. By pre-computing feature attributions and uncertainty measures, the system prepares explanation components in advance, reducing the computational burden during actual decision-making and compliance reporting processes.
Solution Approach 2:
The patent computes explanations for only the most important attributes rather than all possible features. By identifying and explaining only the top-k most influential attributes based on SHAP values, the system achieves regulatory compliance with reduced computational effort, avoiding the intractability of computing exhaustive explanations for all features.
3Measurement precision
If error estimation is performed to improve explanation accuracy, then explanation quality is improved, but confidence interval reduction requires additional computation
Solution Approach 1:
The patent changes the parameter of uncertainty measurement from exact confidence intervals to estimated error bounds. By using statistical error estimation techniques on SHAP values rather than computing exact confidence intervals, the system achieves improved explanation accuracy with reduced computational complexity, trading precise interval bounds for practical error estimates.
Data Source
AI summary
Methods and systems for using a non-linear machine learning model to generate explanations that relate to decisions made by the model and for repairing such explanations in order to improve quality and accuracy of model outputs are provided. The method includes: receiving a data set that corresponds to attributes that pertain to a decision to be made; inputting the data set to a machine learning model; generating a baseline decision that corresponds to an output of the model with respect to data set; computing, based on the baseline decision, an explanation that relates to at least one of the attributes; estimating one or more errors associated with the explanation; and computing, based on the estimated error(s), at least one repair that corresponds to a modification of the explanation, and a cost for repairing the explanation.