Monte Carlo Variable Contribution Estimation in Data Science Models
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Complex data science models find it difficult to interpret how individual input variables contribute to their outputs, especially when considering dependencies among variables, making it challenging to understand which variables have the greatest effect on the model's output.
Innovation Solution
A method is introduced that groups input variables based on their dependencies and uses Monte Carlo sampling to approximate Owen values, allowing for the determination of individual input variable contributions by iteratively selecting random coalitions and computing iteration-specific contribution values, which are then aggregated to determine the overall contribution of each variable.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If data science models are made complex to capture intricate patterns in data, then model predictive accuracy is improved, but interpretability of individual input variable contributions deteriorates
Solution Approach 1:
The patent introduces Owen value computation as an intermediary mechanism that bridges complex data science models and interpretable variable contributions. By using game-theoretic Owen values with variable grouping as a mediator, the system maintains complex model architecture for accuracy while providing interpretable contribution metrics through the Owen value computation process.
2Measurement precision
If exact Owen values are computed to precisely determine variable contributions, then measurement precision of variable importance is improved, but computational resources and time required increase significantly
Solution Approach 1:
The patent applies partial action by computing Owen values only for grouped variables rather than all individual variables, and by using Monte Carlo sampling to compute approximate Owen values rather than exact values. This partial computation approach maintains sufficient precision for interpretability while dramatically reducing computational burden compared to exact Owen value computation for all variables.
Solution Approach 2:
The patent changes the computational parameters by introducing a grouping structure that aggregates variables into meaningful groups, and by using Monte Carlo sampling with a controlled number of iterations. These parameter changes transform the computation from an intractable exact calculation to a manageable approximate calculation that balances precision and efficiency.
3Measurement precision
If all possible variable coalitions are considered in Owen value computation, then accuracy of contribution measurement is improved, but device complexity and computational burden increase exponentially
Solution Approach 1:
The patent segments the set of all variables into smaller grouped variables based on dependencies and relationships. By grouping related variables together, the patent reduces the number of coalitions that need to be evaluated in Owen value computation, transforming an exponentially complex problem into a manageable one while preserving the essential contribution information.
Data Source
AI summary
A computing platform is configured to (i) train a data science model object to receive an input data record including a set of input variables and output a score for the input data record, (ii) arrange the input variables into variable groups based on dependencies between input variables, (iii) identify an input data record to be scored by the model object, (iv) for each respective variable group, iterate the following: (a) identify a sample historical data record from a set of historical data records, (b) select a random group coalition, (c) select a random variable coalition within respective variable group, and (d) use the input data record, the sample historical data record, the randomly-selected group coalition, and the randomly-selected variable coalition to compute an iteration-specific contribution value, and (v) for each respective input variable, aggregate the iteration-specific contribution values and thereby determine an aggregated contribution value for the respective input variable.