Machine Learning Explanation System with Parameter Segmentation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current machine learning systems, particularly in applications like loan examinations, fail to provide clear and actionable explanations for their decisions, making it difficult for users to understand and accept the reasons behind their outcomes, as existing methods focus on extracting explanatory variables rather than providing user-friendly explanations.
Innovation Solution
A system that allows users to set conditions and weights for input parameters, enabling the generation of explanations by changing specific parameters to demonstrate how the outcome would differ, thereby providing actionable insights into the decision-making process.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Loss of information
If machine learning systems provide detailed explanations of all parameters affecting the estimated result, then the completeness of information is improved, but the complexity of the explanation increases making it difficult for users to understand
Solution Approach 1:
The system segments the explanation information by separating parameters that can be changed from parameters that cannot be changed. This is achieved by identifying modifiable parameters (those within user control) and immutable parameters (those outside user control), presenting them in distinct sections of the explanation interface.
Solution Approach 2:
The system extracts and highlights only the most relevant and actionable parameters from the complete set of parameters affecting the estimated result. By identifying and presenting key parameters that have the greatest impact and are modifiable, the system reduces information overload while maintaining explanatory completeness.
2Measurement precision
If machine learning systems provide explanations with specific parameter changes (e.g., age reduction), then the precision of the explanation is improved, but the feasibility of the explanation deteriorates when the changes are unrealistic
Solution Approach 1:
The system dynamically adjusts the explanation based on the modifiability of parameters. For immutable parameters like age, the system presents them as informational context rather than actionable recommendations. For modifiable parameters like income or employment status, the system provides specific, feasible change recommendations that users can actually implement.
Solution Approach 2:
The system applies different qualities of explanation to different parameters based on their modifiability characteristics. Modifiable parameters receive actionable, prescriptive explanations with specific change recommendations, while immutable parameters receive descriptive explanations that provide context without implying feasibility of change.
3Loss of information
If machine learning systems extract multiple explanatory variables contributing to the objective variable, then the comprehensiveness of the analysis is improved, but the usability of the explanation deteriorates due to information overload
Solution Approach 1:
The system extracts and prioritizes the most significant parameters from the complete set of explanatory variables. By identifying key parameters with the greatest impact on the estimated result and filtering out less significant ones, the system maintains analytical comprehensiveness while improving usability through focused, actionable information.
Solution Approach 2:
The system transforms the complete set of explanatory variables into a prioritized hierarchy based on their impact magnitude and modifiability. Parameters are ranked and presented in order of importance, with the most significant and actionable parameters appearing first, allowing users to focus on the most relevant factors.
Data Source
AI summary
A system provides an estimated result of an input including a plurality of parameters using a learning model, and the system is configured to provide information necessary to change an estimated result with regard to one parameter among the plurality of parameters, and set a condition for information with regard to at least one parameter among the plurality of parameters, wherein the condition is a condition for not providing some information regarding a parameter which is a target of the set condition as the information necessary to change the estimated result.