Feature Selection Using Knowledge Graph Superordinate Concepts
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing feature selection methods in explainable AI do not necessarily improve the interpretability of model outputs, even when evaluation indices like AIC are enhanced, as they often select redundant features from large datasets.
Innovation Solution
A feature selection method that identifies and selects features based on superordinate concepts within a knowledge graph, using a feature selection device that constructs training data, specifies superordinate concepts, and determines their inclusion based on hypotheses satisfying certain conditions related to the objective variable, thereby improving model interpretability.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If feature selection is performed using evaluation indices like AIC from large datasets, then model accuracy may be improved, but interpretability of model outputs deteriorates due to selection of redundant features
Solution Approach 1:
The patent segments the feature selection process into two distinct phases: (1) initial feature selection using evaluation indices like AIC to identify potentially useful features, and (2) subsequent filtering using a knowledge graph to remove redundant features. This segmentation allows the system to first optimize for accuracy and then for interpretability, resolving the contradiction between these two objectives.
Solution Approach 2:
The patent introduces a knowledge graph as an intermediary structure that encodes semantic relationships between features. This knowledge graph acts as a mediator between the statistically-selected features and the final interpretable feature set, using pre-defined semantic relationships to identify and eliminate redundant features while preserving those that provide genuine explanatory value.
2Reliability
If multiple features from subordinate concepts are selected to satisfy statistical conditions, then model performance improves, but explanation redundancy increases
Solution Approach 1:
The patent merges multiple subordinate concept features that share the same superordinate concept into a single representative feature. When multiple features from subordinate concepts satisfy statistical conditions, the system combines them by selecting one feature at the superordinate concept level, thereby maintaining model performance while eliminating redundant explanations that would arise from selecting multiple features representing the same underlying concept.
Solution Approach 2:
Instead of the conventional approach of selecting features bottom-up from specific features, the patent inverts the selection logic by working top-down through the knowledge graph hierarchy. It identifies superordinate concepts first, then selectively includes or excludes subordinate features based on whether their parent superordinate concept is already represented, thereby reducing redundancy while preserving performance.
3Ease of manufacture
If features are selected based on statistical evaluation without considering semantic relationships, then selection process is simple, but interpretability of selected features deteriorates
Solution Approach 1:
The patent performs preliminary action by pre-construcing a knowledge graph that encodes semantic relationships between features before the actual feature selection process. This pre-processing step creates a reusable semantic framework that can be applied during feature selection to automatically identify and eliminate redundant features, thereby maintaining interpretability without significantly complicating the selection process itself.
Data Source
AI summary
A non-transitory computer-readable storage medium storing a feature selection program that causes at least one computer to execute a process, the process includes specifying a feature of a superordinate concept that has a feature included in a feature set as a subordinate concept; and selecting the feature of the superordinate concept as a feature to be added to the feature set when a plurality of hypotheses each represented by a combination of features that include the feature of the subordinate concept satisfies a certain condition based on an objective variable, features of the subordinate concept being different from each other.