xLIME Feature Influencer Identification for ML Models
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Machine learned predictive models lack the ability to identify feature contributors and influencers at the individual sample level, providing only global insights into important features, which limits their usefulness in targeted sales predictions and other applications.
Innovation Solution
The implementation of an advanced model interpretation method called xLIME, which uses localized stratified sampling and piecewise linear regression to provide sample-level feature reasoning and distinguish between top positive and negative feature influencers.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Loss of information
If machine learned models provide global-level feature importance only, then model simplicity is maintained, but sample-specific interpretability is lost
Solution Approach 1:
The patent segments the global feature importance analysis into sample-specific local analyses. By dividing the overall model interpretation task into individual sample-level explanations, the system provides targeted feature importance for each sample without requiring complete redesign of the model architecture, thus reducing information loss while managing complexity through localized processing.
Solution Approach 2:
The patent introduces an intermediary explanation layer between the model predictions and user interpretations. This intermediary layer (the explanation generation module) translates complex model decisions into human-readable feature importance statements for each sample, bridging the gap between model complexity and user needs without requiring users to understand the underlying model complexity.
2Productivity
If models provide only global feature importance, then computational resources are conserved, but targeted sales predictions become difficult
Solution Approach 1:
The patent applies partial action by providing feature importance explanations only for the subset of samples that require targeted analysis (e.g., high-value leads for sales teams) rather than computing explanations for all samples. This selective approach maintains productivity for targeted predictions while conserving computational resources by avoiding unnecessary explanations for samples that don't require deep analysis.
3Loss of information
If feature importance is provided at global level only, then model simplicity is preserved, but actionable insights for specific companies are lost
Solution Approach 1:
The patent implements local quality by providing feature importance explanations that are customized to each sample's specific characteristics. Instead of a uniform global explanation, the system generates sample-specific explanations that highlight which features mattered most for each individual prediction, making the insights actionable and relevant to specific companies or samples while maintaining ease of operation through automated explanation generation.
Data Source
AI summary
In an example, for each feature of one or more features of a target sample data, feature values for one or more pseudo-samples are generated using, localized stratified sampling. The one or more pseudo-samples are fed into the trained machine learned model to obtain their prediction values. A piecewise linear regression model is trained using the one or more pseudo-samples and their prediction values, the piecewise linear regression model having two coefficients for each feature, a first coefficient describing prediction change when a corresponding feature value is increased and a second coefficient describing prediction change when a corresponding feature value is decreased. A top positive feature influencer is identified based on a feature of the one or more features of the target sample having a greatest magnitude of positive first coefficient or greatest magnitude of negative second coefficient. A top negative feature influencer is identified based on a feature of the one or more features of the target sample having a greatest magnitude of negative first coefficient or greatest magnitude of positive second coefficient. A top feature contributor is identified based on a feature of the one or more features of the target sample having a greatest magnitude of a combination of second coefficient and feature value in the target sample data.