Feature Influence Quantification in Ensemble Decision Tree Models
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Complex data science models are difficult to interpret, making it challenging to determine which features significantly influence the model's output.
Innovation Solution
A method is disclosed to quantify the influence of specific features on the output of a trained data science model by partitioning features into global groups, using an ensemble of symmetric decision trees, and computing individual and overall contribution values.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If complex data science models are used to derive insights, then prediction accuracy is improved, but interpretability of feature influence deteriorates
Solution Approach 1:
The patent segments the complex model analysis into two distinct components: (1) the prediction function that generates outputs, and (2) the attribution function that separately quantifies feature influences. This segmentation allows the model to maintain its complex predictive capabilities while providing simplified, interpretable feature contribution measurements through the separate attribution mechanism.
Solution Approach 2:
The patent introduces an intermediary attribution function that acts as a mediator between the complex prediction model and the user seeking interpretation. This intermediary component takes the model's internal feature interactions as input and transforms them into human-interpretable contribution values, without requiring changes to the underlying complex prediction model.
2Adaptability or versatility
If complex data science models are used, then insight derivation capability is improved, but ease of interpreting feature effects deteriorates
Solution Approach 1:
The patent implements a self-service mechanism where the attribution function automatically quantifies feature influences without requiring manual analysis or external tools. The system serves its own interpretation needs by internally computing contribution values based on feature perturbations, eliminating the need for users to manually interpret complex model behaviors.
Solution Approach 2:
The patent utilizes parameter changes by systematically perturbing input feature values and observing the resulting changes in model outputs. By measuring how sensitive the prediction is to changes in each feature parameter, the system derives interpretable contribution values that reflect the relative importance of each feature without simplifying the underlying complex model.
3Measurement precision
If traditional feature analysis methods are used on complex models, then computational accuracy is maintained, but computational efficiency deteriorates
Solution Approach 1:
The patent applies preliminary action by pre-computing and storing attribution information in lookup tables during model training or preprocessing. When making predictions, the system can quickly retrieve pre-calculated feature contribution values instead of performing computationally intensive analysis in real-time, significantly improving computational efficiency while maintaining accuracy.
Solution Approach 2:
The patent creates simplified copies of the attribution relationships in the form of lookup tables that map input features to their contribution values. These copied representations allow the system to answer feature influence queries efficiently without repeatedly executing the full complex analysis, trading off some computational resources during table creation for rapid query response later.
Data Source
AI summary
A computing platform is configured to (i) receive a request to compute a score for an input data record; (ii) partition a set of features into global feature groups; (iii) input a group of actual parameters associated with the features into a trained data science model comprising an ensemble of decision trees; (iv) for each tree in the ensemble, identify a respective leaf based on a comparison of the actual parameters to a series of splitting conditions for the respective leaf and determining respective individual contribution values for features for the respective leaf based on local feature groups corresponding to the global feature groups; (v) compute a respective overall feature contribution value for each individual feature; (vi) compute the score for the input data record; (vii) identify a reason code for the score; and (viii) transmit the score and the reason code in response to the request.