Counterfactual Sample Aggregation for Sub-Population ML Explanations
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing counterfactual explanation techniques in explainable machine learning are limited to single predictions, fail to provide insights for sub-populations, are labor-intensive for textual explanations, and struggle with model updates and user preferences, leading to inefficiencies and infeasible recommendations.
Innovation Solution
A computing system aggregates multiple counterfactual samples to generate explanations for sub-populations, uses clustering models to assign samples and determine explanations, alternates between loss functions for model training to respect previous recommendations, and applies user preferences to control feature changes in counterfactual sample generation.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Loss of information
If existing counterfactual explanation techniques are used, then a single prediction can be explained, but insights for sub-populations cannot be provided
Solution Approach 1:
The patent segments the population into sub-populations using clustering algorithms (e.g., K-means, DBSCAN) based on feature similarities. Each cluster represents a distinct sub-population that can be independently analyzed and explained, allowing the system to provide targeted explanations for different groups rather than treating all predictions uniformly.
Solution Approach 2:
The patent combines multiple counterfactual explanations across different clusters to form comprehensive sub-population level explanations. By aggregating individual counterfactuals from clustered samples, the system generates explanations that represent entire sub-populations, thereby providing broader insight coverage without analyzing every single prediction individually.
2Measurement precision
If textual explanations are created for each individual counterfactual sample, then detailed explanations are provided, but the process becomes labor intensive and inefficient
Solution Approach 1:
The patent generates textual explanations by clustering similar counterfactual samples and creating representative explanations for each cluster. Instead of crafting detailed text for every individual counterfactual, the system creates template explanations that can be copied and adapted across multiple similar cases, significantly improving efficiency while maintaining sufficient detail through the clustering-based grouping.
3Reliability
If machine learning models are updated or replaced, then model performance may improve, but previously generated counterfactual recommendations become obsolete
Solution Approach 1:
The patent incorporates counterfactual stability constraints into the model training process, penalizing excessive changes in predictions when model parameters are updated. By adding regularization terms to the loss function that measure prediction drift, the system maintains greater consistency in recommendations across model versions while still allowing necessary performance improvements.
4Manufacturing precision
If counterfactual sample generation modifies feature values to change model output, then model output changes are achieved, but user preferences and constraints may be violated
Solution Approach 1:
The patent implements dynamic counterfactual generation that adapts to user preferences and constraints by making the feature modification process adjustable. The system allows users to specify which features can be modified, sets bounds on modification magnitudes, and prioritizes certain features based on user preferences. This dynamic approach enables the system to achieve desired output changes while respecting user-defined constraints and preferences.
Data Source
AI summary
In some aspects, a computing system may aggregating multiple counterfactual samples so that machine learning explanations can be generated for sub-populations. In addition, methods and systems described herein use machine learning and counterfactual samples to determine text to use in an explanation for a model's prediction. A computing system may also train machine learning models to not only determine whether a request to perform an action should be accepted, but also to generate output that is consistent with output generated by previous machine learning models. Further, a computing system may generate counterfactual samples based on user preferences. A computing system may obtain preferences and then apply a penalty or adjustment parameter such that when a counterfactual sample is created, the computing system is forced to change one or more features indicated by the preferences to create the counterfactual sample.