Knowledge-Graph Feature Engineering for Interpretable Models
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing feature selection methods for machine learning algorithms are challenging for non-experts, requiring manual effort and intuition, and lack automation for generating features that are both statistically important and interpretable by domain experts.
Innovation Solution
A scalable solution that automates feature engineering using a knowledge graph and reinforcement learning to derive interpretable and statistically-important features, balancing model performance and interpretability through a bi-objective optimization process.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Loss of information
If manual feature selection is performed by data scientists using intuition and domain knowledge, then feature interpretability is improved, but productivity deteriorates due to inordinate time and expense
Solution Approach 1:
The patent introduces an automated feature engineering system that acts as an intermediary between raw data and machine learning models. This system uses a knowledge graph containing domain knowledge and rules to automatically generate interpretable features, eliminating the need for manual feature selection while preserving interpretability through structured domain knowledge representation.
Solution Approach 2:
The system enables self-service feature engineering by automatically generating features using a knowledge graph and rule-based transformations. The automated feature engineering component performs feature selection and transformation without human intervention, allowing the system to serve itself in the feature engineering process while maintaining interpretability through domain knowledge integration.
2Productivity
If automated feature selection systems are used to increase productivity, then feature interpretability deteriorates because existing systems focus only on statistical importance
Solution Approach 1:
The patent creates a composite feature engineering system that combines multiple components: a knowledge graph containing domain knowledge, rule-based transformation systems, and automated feature generation. This composite approach integrates both statistical importance (through automated evaluation) and interpretability (through domain knowledge integration), producing features that satisfy both criteria simultaneously.
Solution Approach 2:
The knowledge graph serves multiple functions: it stores domain knowledge, provides rules for feature transformation, ensures feature interpretability, and guides automated feature generation. This multi-functional component enables the system to achieve both automation and interpretability by universally applying domain knowledge across the feature engineering process.
3Reliability
If the number of possible features is increased to improve model performance, then device complexity deteriorates due to the unlimited number of transformable features
Solution Approach 1:
The patent applies local quality by generating features with specific properties tailored to the domain and problem at hand. Instead of generating all possible features uniformly, the system uses domain-specific rules and knowledge graph constraints to generate only relevant features with appropriate transformations, reducing complexity while maintaining performance.
Solution Approach 2:
The system changes parameters by dynamically adjusting feature transformations based on domain knowledge and performance feedback. The automated feature engineering component modifies feature parameters (transformations, combinations, selections) to optimize model performance without exhaustively exploring all possible feature spaces, thereby managing complexity.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
Systems and methods include generation of a first set of features based on a second set of features and a learning network, determination of an interpretability value for each of the first set of features, determination of a performance of a model trained using the first set of features, determination of a reward based on the performance and the interpretability values, and generation of a third set of features based on the first set of features, the learning network and the reward.