Class-Specific Predictor Importance for Interpretable Tree Models
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing tree-based machine learning models do not provide insights into the importance of specific predictor indicators for specific classes or categories, limiting their interpretability and effectiveness in applications like medical diagnosis and business analytics.
Innovation Solution
A system and method to determine class-specific predictor importance by generating a tree structure with nodes and class-specific predictor indicators, allowing for graphical visualization and quantification of the importance of predictor indicators for specific classes.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If tree-based machine learning models are used for predictions, then the models can process both continuous and categorical predictors without normalization and handle non-linear relationships, but the models do not provide insights into the importance of specific predictor indicators for specific classes
Solution Approach 1:
The patent segments the overall predictor importance into class-specific predictor importances by introducing class-specific predictor indicators at each node of the decision tree. This segmentation allows the model to provide detailed insights about which predictors are important for each specific class while maintaining the overall tree structure for predictions.
Solution Approach 2:
The patent applies local quality by assigning different importance measurements to different parts of the tree structure. Specifically, it calculates purity predictor indicator frequencies and term predictor indicator frequencies at each node, allowing different regions of the tree to have different predictor importance characteristics based on their local class distributions.
2Ease of operation
If the tree structure is visualized with all nodes and paths, then the model predictions become more intuitive and explainable, but the complexity of the graphical representation increases
Solution Approach 1:
The patent extracts and highlights only the relevant paths and nodes that are important for explaining specific class predictions. By focusing on class-specific predictor indicators and their frequencies, the system extracts the essential explanatory information from the full tree structure without requiring visualization of every node and path, thus reducing graphical complexity while maintaining explainability.
Solution Approach 2:
The patent adds a new dimension to the tree visualization by incorporating class-specific predictor indicator frequencies and purity measurements. This additional dimensional information allows the graphical representation to convey importance metrics without increasing the basic tree structure complexity, as the new information is layered onto the existing structure rather than requiring separate visualizations.
Data Source
AI summary
Determining a class-specific predictor indicator importance can include generating, using machine learning, a prediction model having a tree structure including nodes, class label subgroups, and class-specific predictor indicators. A class-specific predictor indicator importance can be generated for a selected class-specific predictor indicator by merging a term predictor indicator frequency and a purity predictor indicator frequency at each instance of the selected class-specific predictor indicator within one or more nodes of the tree structure. A representation of the tree structure displayed in a graphical user interface can depict each path between a root node of the tree structure and a leaf node mapped to the selected class label subgroup having a node containing an instance of the selected class-specific predictor indicator in a manner that contrasts with other paths between the root node and other leaf nodes of the tree structure.