Shapley Value Feature Contribution Analysis System
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Modern predictive models lack the ability to interpret their prediction results, making it difficult for users to understand how inputs contribute to output values, which reduces confidence in model performance and hinders insights into the modeled system.
Innovation Solution
A computing device executes instructions to determine the contribution of each feature to a predicted value using a selection vector matrix, feature computation dataset, and Shapley estimate values, providing a more interpretable explanation of predictive model outputs.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional predictive models are used, then prediction accuracy can be achieved, but interpretability of prediction results is lost
Solution Approach 1:
The patent segments the prediction process into multiple components by introducing selection vector matrices that divide the feature space into different subsets. Each selection vector represents a specific feature combination, allowing the model to break down the complex prediction into interpretable feature contributions while maintaining prediction accuracy through systematic evaluation of different feature segments.
Solution Approach 2:
The patent introduces selection vector matrices as an intermediary between the input features and the prediction output. These matrices serve as a mediator that captures the relationship between features and predictions in a structured, interpretable format, enabling users to understand which features contribute to predictions without sacrificing the model's predictive power.
2Loss of information
If feature contribution analysis is performed using traditional methods, then interpretability is improved, but computational complexity increases
Solution Approach 1:
The patent performs preliminary action by pre-defining selection vector matrices with all possible feature combinations before the actual prediction analysis. This preprocessing step organizes the computational work in advance, allowing the model to efficiently query pre-computed feature contributions rather than calculating them from scratch during prediction, thereby reducing real-time computational complexity.
Solution Approach 2:
The patent applies partial action by allowing users to select specific selection vectors from the matrix based on their needs. Instead of computing all possible feature contributions, users can focus on particular feature subsets relevant to their analysis, reducing the computational burden while maintaining interpretability for the most important features.
3Loss of information
If comprehensive feature analysis is conducted, then model interpretability is enhanced, but computation time increases
Solution Approach 1:
The patent computes feature contributions using pre-defined selection vector matrices during the model training or preprocessing phase. By performing this computationally intensive analysis in advance, the results are stored and can be quickly retrieved during deployment, significantly reducing the computation time required for interpretability analysis when the model is actually used.
Solution Approach 2:
The patent creates a copy of the feature space representation in the form of selection vector matrices that encode all possible feature combinations. This copied structure allows the model to quickly reference pre-computed feature contributions without re-running complex calculations, enabling fast interpretation during inference while maintaining comprehensive analysis capabilities.
Data Source
AI summary
Systems and methods for machine learning, models, and related explainability and interpretability are provided. A computing device determines a contribution of a feature to a predicted value. A feature computation dataset is defined based on a selected next selection vector. A prediction value is computed for each observation vector included in the feature computation dataset using a trained predictive model. An expected value is computed for the selected next selection vector based on the prediction values. The feature computation dataset is at least a partial copy of a training dataset with each variable value replaced in each observation vector included in the feature computation dataset based on the selected next selection vector. Each replaced variable value is replaced with a value included in a predefined query for a respective variable. A Shapley estimate value is computed for each variable.