Kernel Model Expansion for Interpretability
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Machine learning models face a trade-off between interpretability and performance, where high-interpretability models have low accuracy and high-performance models like kernel machines are difficult to interpret.
Innovation Solution
A model generation system that uses a kernel function generation unit to define a scalable white-box kernel function via random Fourier features, a model learning unit to learn a linear model, and a component expanding unit to expand the model for improved interpretability and performance.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a high-performance model like kernel machine or deep neural network is used, then prediction accuracy is improved, but interpretability deteriorates due to model complexity
Solution Approach 1:
The patent introduces an intermediary representation layer that connects the complex kernel machine model with interpretable feature components. By defining expanded component representation that includes interaction terms between features, the system mediates between the black-box kernel predictions and human-understandable feature contributions, allowing both high accuracy and interpretability to coexist
Solution Approach 2:
The patent segments the complex kernel model into individual feature components and their interactions. By expanding the kernel representation into sum-of-products form with explicit feature terms, the model breaks down the monolithic complex structure into interpretable segments, each representing a specific feature's contribution or interaction effect
2Device complexity
If a high-interpretability model like additive model is used, then interpretability is improved, but prediction accuracy deteriorates due to model simplicity
Solution Approach 1:
The patent creates a composite model structure that combines the simplicity of additive models with the power of kernel methods. By forming a composite representation that includes both main effect terms and interaction terms in an expanded sum-of-products form, the model achieves the interpretability of additive models while incorporating the predictive power of kernel machines
3Productivity
If kernel machine with powerful model structure is used, then performance is improved, but interpretability deteriorates
Solution Approach 1:
The patent transforms the parameter representation of the kernel model from implicit kernel space to explicit feature space with interaction terms. By changing the parameterization to include expanded component representation with clear feature dependencies, the model maintains its performance while becoming interpretable through explicit parameter definitions
Data Source
AI summary
The kernel function generation unit 81 defines a first kernel function by using two-dimensional feature representation that represents a combination of two features of data. The model learning unit 82 defines a linear model including an inner product of a mapping used in the first kernel function and a first weight and performs learning with the defined linear model. The component expanding unit 83 expands the learned linear model to define expanded component representation that is new component representation of the data. The expansion model generation unit 84 generates an expansion model including an inner product of data by the expanded component representation and a second weight.