Machine Learning Predictor Profiling via Stratified Sampling
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing machine learning models lack effective methods for explaining predictions and profiling influential predictors, particularly due to random sampling issues in local interpretable model-agnostic explanation methods, which require a large number of samples to build a decent predictive model and do not provide sufficient insights into the ranges of varying predictors.
Innovation Solution
An iterative method using a fine-grained sampling strategy to detect and profile influential predictors by determining intervals where predictors can vary without changing the prediction, generating sample cases across these intervals, and iteratively refining the intervals until a stopping rule is satisfied, allowing for the construction of a predictive model that provides insights into the ranges of influential predictors.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If random sampling is used in local interpretable model-agnostic explanation methods, then a large number of samples are required to build a decent predictive model, but this increases the sampling effort and time required
Solution Approach 1:
The patent changes the sampling parameters from completely random sampling to stratified sampling based on predictor ranges. By dividing the predictor space into intervals and sampling systematically from each interval, the method achieves better prediction explanation accuracy without requiring a large number of samples, thus reducing sampling time.
Solution Approach 2:
The patent performs preliminary action by first determining the range of each predictor and dividing it into intervals before sampling. This pre-structuring of the sampling space ensures that samples are distributed effectively across all relevant predictor variations, improving explanation accuracy while minimizing the total number of samples needed.
2Loss of information
If random sampling is used to build predictive models for explanation, then sufficient insights into predictor ranges are not provided, but this limits the interpretability of model predictions
Solution Approach 1:
The patent applies segmentation by dividing the continuous predictor range into discrete intervals. This segmentation allows the method to capture and provide insights into how predictions change across different predictor ranges, thereby reducing information loss about predictor behavior while maintaining a structured and manageable explanation framework.
3Reliability
If a large number of samples are collected through random sampling, then a decent predictive model can be built, but this increases the computational resources and sampling effort required
Solution Approach 1:
The patent changes the sampling approach from random to stratified sampling across predictor intervals. This parameter change ensures that samples are distributed to cover all relevant predictor variations, improving predictive model reliability without requiring a large number of samples, thus maintaining high sampling efficiency.
Data Source
AI summary
In an approach for detecting and profiling influential predictors in a machine learning model, a processor receives a case as an input to a pre-trained model for a prediction, the prediction being based on a plurality of predictors. A processor determines a range for a predictor where the predictor varies without changing the prediction. A processor generates a data set by sampling data across the determined range of the predictor. A processor builds a predictive model with the generated data set. A processor outputs a prediction result including an insight about the range of the predictor.