Linear Model Classification for Interpretable Machine Learning
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing machine learning models, particularly black-box types, lack interpretability, leading to user dissatisfaction due to unclear determination processes, and interactive approaches to select important explanatory variables are costly in terms of calculation and user engagement.
Innovation Solution
Classifying linear models into equivalence classes based on identical variables and coefficient encoding, then asking questions to users about the importance of explanatory variables to reduce the number of candidate groups, thereby narrowing down models and reducing calculation costs.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a black-box machine learning model is used to achieve high accuracy, then prediction performance is improved, but interpretability deteriorates making user acceptance difficult
Solution Approach 1:
The patent segments the machine learning model into two distinct components: a black-box model for high-accuracy predictions and a white-box model for interpretable explanations. This segmentation allows each component to fulfill its specific function - the black-box model handles complex pattern recognition while the white-box model provides human-understandable reasoning, thereby resolving the contradiction between accuracy and interpretability
Solution Approach 2:
The patent introduces a white-box model as an intermediary between the black-box model and the user. This intermediary translates the opaque predictions into interpretable explanations that users can understand and trust, maintaining the high accuracy of the black-box model while adding the necessary layer of interpretability for user acceptance
2Loss of information
If a white-box machine learning model is used to ensure interpretability, then human-interpretable determination is achieved, but prediction accuracy deteriorates
Solution Approach 1:
The patent divides the modeling task into two specialized segments: a black-box model optimized for prediction accuracy and a white-box model optimized for interpretability. Each model type is used for its strength - the black-box model captures complex non-linear relationships while the white-box model provides transparent decision logic, eliminating the need to choose one over the other
Solution Approach 2:
The patent merges the strengths of both black-box and white-box models into a unified system. The black-box model generates predictions while the white-box model provides explanations, creating a combined approach that achieves both high accuracy and interpretability simultaneously, rather than treating them as mutually exclusive options
3Adaptability or versatility
If an interactive approach is used to generate acceptable machine learning models, then user satisfaction is improved, but calculation cost and user engagement time increase
Solution Approach 1:
The patent performs preliminary action by pre-generating multiple candidate models with different characteristics before user interaction. This allows the system to present users with curated options rather than requiring them to explore the entire model space, significantly reducing the time and effort needed for model selection while maintaining high user satisfaction
Solution Approach 2:
The patent applies local quality by tailoring model explanations to match specific user preferences and domain knowledge. Rather than providing generic explanations, the system adapts the interpretability characteristics to local user needs, making the interaction more efficient and satisfying while reducing the overall engagement time required
Data Source
AI summary
An information processing device classifies a plurality of linear models, each of which includes one or more variables, into a plurality of groups in such a way that the linear models which include identical variables included in each of the plurality of linear models and which have identical coefficient encoding with respect to the variables are grouped in the same group, outputs a first question used in deciding degree of importance of each explanatory variable included in training data which is used in training of the plurality of linear models, and, decides on an explanatory variable about which a second question is to be asked, when a linear model in which the degree of importance is reflected is to be selected from the plurality of linear models, based on extent of decrease in number of target groups for selection according to an answer to the first question.