Automated Classification Explanation via Feature Adjustment
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current methods for assessing and understanding classification models, especially complex ones like neural networks, lack visibility into the features that influence classification outcomes and how changes in these features affect results, making it difficult to explain or modify model performance.
Innovation Solution
A system and method that automatically generates descriptions for classification results by adjusting input features, reclassifying the inputs, and training a machine-learning model to identify adjustments that would change the classification outcome, providing explanations and actionable insights into the model's performance.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If manual approaches are used to analyze classification models by plotting distributions and hand-selecting thresholds, then limited understanding into simple classification models can be obtained, but these approaches are infeasible for complex classification models such as neural networks
Solution Approach 1:
The system enables classification models to explain their own decisions automatically. The explanation generator analyzes the model's internal structure and generates feature importance explanations without requiring external manual analysis, making the approach self-serving and automatically adaptable to different model types including complex neural networks.
Solution Approach 2:
The patent replaces manual mechanical analysis methods (plotting distributions, hand-selecting thresholds) with an automated computational explanation generation system. This substitution enables the analysis to scale to complex models like neural networks that were previously infeasible to analyze manually.
2Productivity
If automated methods are used for assessing classification model performance, then speed and accuracy assessment is improved, but visibility into what features of input data are most important is lost
Solution Approach 1:
The system segments the classification process into distinct analyzable components by identifying and ranking individual features and their contributions to the classification decision. This segmentation allows automated assessment to maintain speed while preserving visibility into feature importance through structured explanation generation.
Solution Approach 2:
The explanation generation system provides feedback about feature importance and decision rationale back to the user. This feedback loop maintains information visibility by automatically generating and returning explanations that highlight which features were most influential in the classification outcome.
3Productivity
If no explanation generation system is used, then classification results are produced quickly, but understanding into how classification models generate output is minimal
Solution Approach 1:
The system generates explanations selectively for specific classification results rather than for all outputs. This partial action approach maintains productivity by avoiding unnecessary explanation generation while still providing understanding where needed, balancing speed with information availability.
Solution Approach 2:
The explanation generator acts as an intermediary between the classification model and the user. It translates the model's internal decision-making process into human-understandable feature importance explanations, preserving classification rationale without interfering with the model's fast output generation.
Data Source
AI summary
A method for generating descriptions of classification results, comprising classifying original inputs in a data set using a classifier. For one or more of the plurality of original inputs, apply an adjustment to one or more features of the original input to produce a respective adjusted input, and classify the respective adjusted input using the classifier. For one or more of the adjusted inputs, record a corresponding original input into a training set, and record information associated with the adjusted input into the training set. The method further comprises training a model based on the training set, such that the model is configured to receive an original model input and identify a hypothetical adjustment that would produce an adjusted model input that would be assigned an adjusted model classification with a different class than a class assigned to the original model input.