Adaptive Sampling for Shapley Value Computation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current methods for computing global feature importance in machine learning models, such as Shapley values, are inefficient due to the need to evaluate numerous permutations, leading to high computation time and potential quality issues with sampling techniques that do not guarantee explanation quality.
Innovation Solution
An adaptive sampling technique that iteratively samples and evaluates subsets of data, comparing global explanations until a convergence threshold is met, allowing for exponential growth in sample size to ensure precision and quality of feature importance values.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a large set of local explanations is evaluated to compute global feature importance using Shapley values, then the quality of feature importance values is improved, but the computation time increases significantly
Solution Approach 1:
The patent applies partial action by evaluating only a carefully selected subset of local explanations rather than all possible permutations. The adaptive sampling technique identifies and evaluates only the most informative samples, achieving sufficient explanation quality without computing the complete set of O(2^n) permutations that would be required for exact Shapley values.
Solution Approach 2:
The patent changes the parameter of sample size from fixed to adaptive. The adaptive sampling technique dynamically adjusts the number of samples evaluated based on the observed convergence of feature importance values. This allows the computation to stop early when sufficient precision is achieved, significantly reducing computation time while maintaining quality.
2Productivity
If a fixed number of samples is randomly sampled and computation is parallelized, then the computation time is reduced, but the quality of explanation is not guaranteed
Solution Approach 1:
The patent implements feedback through the adaptive sampling mechanism that continuously monitors the convergence of feature importance values. After each batch of samples is evaluated, the algorithm checks whether the feature importance estimates have stabilized within a acceptable margin. This feedback loop ensures that computation stops only when sufficient quality is achieved, providing reliability guarantees that fixed-sample methods cannot offer.
Solution Approach 2:
The patent transforms the static fixed-sample approach into a dynamic adaptive sampling process. The sample size and evaluation criteria are adjusted dynamically based on the observed properties of the data and the convergence behavior of the Shapley value estimates. This dynamic adaptation allows the system to achieve both high productivity and reliable quality guarantees.
3Measurement precision
If the sampled dataset is increased to improve explanation quality, then the precision of feature importance values is improved, but the computation time increases without bound
Solution Approach 1:
The patent applies partial action by evaluating only the necessary subset of samples required to achieve convergence. The adaptive sampling technique identifies when sufficient precision has been reached and stops sampling, avoiding the excessive computation that would result from evaluating all possible permutations or using overly large fixed samples.
Solution Approach 2:
The patent replaces the mechanical approach of systematically evaluating all permutations with an intelligent sampling strategy. Instead of mechanically computing every possible coalition, the algorithm uses heuristics and convergence criteria to substitute selective sampling, dramatically reducing computation time while maintaining precision.
Data Source
AI summary
Techniques for computing global feature explanations using adaptive sampling are provided. In one technique, first and second samples from an dataset are identified. A first set of feature importance values (FIVs) is generated based on the first sample and a machine-learned model. A second set of FIVs is generated based on the second sample and the model. If a result of a comparison between the first and second FIV sets does not satisfy criteria, then: (i) an aggregated set is generated based on the last two FIV sets; (ii) a new sample that is double the size of a previous sample is identified from the dataset; (iii) a current FIV set is generated based on the new sample and the model; (iv) determine whether a result of a comparison between the current and aggregated FIV sets satisfies criteria; repeating (i)-(iv) until the result of the last comparison satisfies the criteria.