Decision Tree Path Explanations With Redundancy-Free Feature Sets
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing decision tree models face challenges in efficiently determining succinct and redundancy-free explanations for predictions, particularly in high-risk applications, due to the NP-hard nature of achieving optimal succinctness and potential arbitrary redundancy in path-based explanations.
Innovation Solution
A method and device for processing decision data in decision trees that identify a universal set of irrelevant features by applying inheritance processing criteria, using a MaxSAT solver to determine a broadest universal set of features irrelevant to explain the decision path, thereby reducing redundancy and achieving polynomial runtime.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of operation
If path-based explanations are used in decision trees, then explainability is improved, but redundancy in explanations occurs and computing optimal succinctness becomes NP-hard
Solution Approach 1:
The patent extracts and removes redundant features from the explanation set by identifying features that do not contribute to distinguishing the decision path. Through inheritance processing, the method separates relevant features (those on the decision path) from irrelevant features, producing a minimal explanation set that maintains explainability while eliminating redundancy.
Solution Approach 2:
The patent applies partial action by processing only the necessary portions of the decision tree related to the specific decision path, rather than analyzing the entire tree. The inheritance processing function selectively evaluates edges and nodes relevant to the current path, reducing computational effort while maintaining explanation quality.
2Productivity
If inheritance processing is applied to determine universal sets of irrelevant features, then redundancy is reduced and runtime becomes polynomial, but the complexity of feature set processing increases
Solution Approach 1:
The patent performs preliminary action by pre-processing the decision tree structure to identify and store inheritance relationships between edges and nodes. This preparation allows the main inheritance processing function to efficiently determine universal sets of irrelevant features without re-evaluating the entire tree structure, achieving polynomial runtime while managing processing complexity.
Solution Approach 2:
The patent applies universality by creating a general inheritance processing function that can handle any decision path and feature set. The function universally determines irrelevant features for any given path, making the approach applicable to various decision tree structures without requiring path-specific custom processing logic.
3Loss of information
If all features are considered in explaining a decision path, then completeness of explanation is improved, but redundancy and processing time increase
Solution Approach 1:
The patent extracts and removes redundant features from the complete feature set by systematically identifying features that do not contribute to the decision path. The inheritance processing function evaluates each feature's relevance and separates it into relevant or irrelevant categories, maintaining explanation completeness while reducing processing time by focusing only on relevant features.
Solution Approach 2:
The patent segments the feature set into distinct categories: features that are relevant to the decision path and features that are irrelevant. This segmentation allows the method to process only the relevant features for explanation generation, reducing processing time while maintaining completeness through the systematic evaluation of all features via inheritance processing.
Data Source
AI summary
Devices and methods for processing decision data including a decision tree and an input path, each node of the decision tree being associated with a respective feature of a feature set. The method includes an inheritance processing, as a function of a current universal set. The inheritance processing includes: determining that the edge verifies a first criterion relative to a consistency of the edge with a corresponding edge of the input path relative to the same feature as those of the current node, and then performing a first sub-inheritance processing as a function of the child node; and/or determining that the edge verifies a second criterion, and then performing a second sub-inheritance processing as a function of the child node and the current universal set, such that the first sub-inheritance processing and the second sub-inheritance processing allows including features in the set for explaining the input path.