Tree Model Explanation via Cached Shapley Attribution
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current methods for explaining tree-based machine learning models using interventional Shapley values are impractical due to high computational complexity, making them inefficient for real-time applications in contexts like credit underwriting, where accurate and rapid explanations are required.
Innovation Solution
The proposed solution involves a model explanation system that utilizes a reference cache module and test cache module to generate and store pre-calculated partial attribution values, reducing computational complexity by focusing on feature-by-feature and leaf-wise computations, and employing reference-averaged partial contributions to facilitate efficient Shapley value calculations.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If interventional Shapley values are used to explain tree-based models, then explanation accuracy and reliability are improved, but computational complexity increases making real-time application impractical
Solution Approach 1:
The patent segments the computation of Shapley values by organizing references into groups based on their traversal paths through the decision tree. Instead of computing Shapley values for all references individually, the system divides them into segments (groups) that share common path characteristics, allowing partial attribution values to be computed once per group and reused across multiple references. This segmentation reduces redundant computations while preserving explanation accuracy.
Solution Approach 2:
The patent performs preliminary actions by pre-computing partial attribution values for groups of references before they are needed for actual explanations. The system pre-calculates traversal paths, organizes references into groups, and computes partial Shapley values in advance, storing them for rapid retrieval during real-time explanation generation. This preliminary computation phase separates the heavy computational burden from the real-time explanation phase.
2Measurement precision
If thousands of reference samples are used to accurately capture non-linear and interactive effects, then explanation quality is improved, but computational runtime increases beyond real-time requirements
Solution Approach 1:
The patent merges the computation of partial attribution values across multiple references by identifying and combining references that share common traversal paths through the decision tree. When multiple references traverse the same path or share common sub-paths, the system merges their computation by calculating partial Shapley values once and reusing them across all references in that group. This merging eliminates redundant computations while maintaining the use of thousands of references for accurate explanation quality.
Solution Approach 2:
The patent performs preliminary organization of references into groups based on their traversal paths before computation. By pre-grouping references that share common paths and pre-computing the partial attribution values for these groups, the system prepares the computational structure in advance. This preliminary action allows the system to handle thousands of references efficiently during real-time operation by simply retrieving pre-computed values rather than recalculating them.
3Loss of information
If Shapley values are calculated for each feature and reference combination, then explanation completeness is improved, but computational efficiency deteriorates
Solution Approach 1:
The patent segments the feature-attribution computation by organizing it around decision tree paths rather than individual feature-reference combinations. The system divides the computation into segments corresponding to different traversal paths through the tree, with each segment containing the partial attribution values for features along that path. This segmentation allows the system to compute explanations for all features across all references by processing path segments once and reusing them, maintaining completeness while improving efficiency.
Solution Approach 2:
The patent performs preliminary computation of partial attribution values for each path segment before they are needed for complete explanations. By pre-computing these partial values and organizing them by path segment, the system prepares all necessary components for complete feature-attribution explanations in advance. During real-time operation, the system simply assembles these pre-computed segments to generate complete explanations, maintaining explanation completeness while achieving computational efficiency.
Data Source
AI summary
Systems and cache-based methods for explaining tree-based models using interventional Shapley values are disclosed. With this technology, interventional Shapley values are used to compute attribution values from a leaf-wise approach within tree-based machine learning models. Reference traversal tables and test traversal tables are created and stored for each leaf of a decision tree. Based on the created tables, a subset of traversal permutations and respective subset size are determined on a tree by tree, leaf by leaf and feature by feature basis. For each of the nodes in a traversal path to each of the leaves, partial attribution values are generated, and an attribution for the node is adjusted based on the generated partial attribution values and a multiplier indicated in the reference traversal tables. An output explanation of a score can advantageously be obtained with reduced computational complexity and runtime.