AI Prediction Interpretation Index for User Trust
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current methods for explaining AI prediction basis lack persuasiveness and cannot be applied to AI models that achieve high prediction accuracy based on combinations of input items, rather than similarity of item values.
Innovation Solution
A computer system that calculates and outputs interpretation indices to select and display learning data useful for users to understand AI predictions, using a predictor and index calculation units to generate display information showing the contribution of each feature to the predicted value.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If black-box processing of AI model is accelerated to improve prediction accuracy, then prediction accuracy is improved, but user trust in predicted value deteriorates
Solution Approach 1:
The patent introduces an interpretation index as an intermediary between the AI model's black-box processing and the user. This index translates the internal decision-making process into comprehensible information that maintains prediction accuracy while building user trust through explainability.
Solution Approach 2:
The patent segments the AI prediction process into two distinct parts: the prediction function that generates accurate results and the interpretation index that explains the reasoning. This segmentation allows both high accuracy and transparency to coexist by separating the computational core from the explanation layer.
2Ease of operation
If contribution information is displayed based on similarity of item values, then ease of operation is improved, but applicability to AI models deteriorates
Solution Approach 1:
The patent changes the fundamental parameter for calculating contribution information from similarity of item values to interpretation indices derived from the AI model's actual decision-making process. This parameter change enables the system to work with any AI model regardless of its internal complexity while maintaining ease of operation.
3Device complexity
If only perturbation data calculation methods are used to explain prediction basis, then device complexity is reduced, but persuasiveness of explanation deteriorates
Solution Approach 1:
The patent incorporates feedback from the AI model's actual prediction process into the explanation system. By using interpretation indices that reflect the model's true decision-making logic rather than external perturbation analysis, the system maintains simplicity while dramatically improving persuasiveness.
Data Source
AI summary
There is provided is a computer system that outputs a predicted value of data to be evaluated using a predictor generated using learning data. The computer system includes the predictor; an index calculation unit that calculates an interpretation index of the data to be evaluated; and an extraction unit that selects the learning data useful for a user to interpret the predicted value of the data to be evaluated, wherein index management information for managing an interpretation index of the learning data is stored, the index calculation unit calculates the interpretation index of the data to be evaluated, and the extraction unit calculates a selection index based on the interpretation index of the data to be evaluated and the interpretation index of the learning data, selects the learning data based on the selection index, and outputs display information for presenting information indicating a processing result.