Ensemble Voting Classifier Threshold Adjustment
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing ensemble voting classifiers face challenges in setting voting thresholds to balance sensitivity and specificity, leading to issues with false positive classifications and over-fitting to training data.
Innovation Solution
A system that adjusts the voting thresholds of each classifier based on vote contributions during training, allowing for increased specificity without impacting sensitivity, using a threshold table and iterative adjustment methods to optimize threshold settings.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If individually set voting thresholds are used for each classifier to meet training set requirements, then voting accuracy improves, but the method cannot control both sensitivity and specificity simultaneously
Solution Approach 1:
The patent changes the parameter representation from fixed thresholds to threshold indices that can be adjusted independently for each classifier. This allows systematic control of sensitivity and specificity by modifying threshold values based on desired performance characteristics, resolving the contradiction between accuracy and adaptability.
Solution Approach 2:
The patent introduces dynamic threshold adjustment where thresholds are not fixed but can be modified based on sensitivity-specificity requirements. The system allows flexible tuning of classifier thresholds to meet different operational requirements, transforming a static system into a dynamic one that adapts to various performance needs.
2Stability of the object's composition
If same thresholds are applied to production items as in training, then consistency is maintained, but over-fitting to training data occurs reducing generalization
Solution Approach 1:
The patent performs preliminary threshold optimization on training data to establish threshold indices that achieve desired sensitivity and specificity. By pre-adjusting thresholds based on training set analysis, the system prepares optimal settings that generalize better to production data while maintaining consistency through the threshold index mechanism.
Solution Approach 2:
The patent uses feedback from training set performance analysis to adjust threshold indices. The system analyzes voting patterns and threshold contributions during training, then modifies thresholds accordingly to improve generalization. This feedback loop ensures that thresholds are optimized based on actual performance rather than assumed consistency.
3Productivity
If genetic optimization is applied directly to voting thresholds, then optimization is attempted, but acceptable results are not yielded
Solution Approach 1:
The patent segments the threshold optimization problem by introducing threshold indices that decouple the optimization of individual classifier thresholds from the overall ensemble performance. This segmentation allows systematic adjustment of each classifier's contribution without requiring complex genetic optimization, improving both efficiency and accuracy.
Solution Approach 2:
The patent introduces threshold indices as intermediaries between raw classifier outputs and final voting decisions. These indices serve as a mediating layer that simplifies the optimization process, allowing direct control over classifier contributions without the complexity of direct genetic optimization on threshold values.
4Device complexity
If mean of output scores is used as final classification score, then simple combination is achieved, but sensitivity to calibration distribution differences reduces accuracy
Solution Approach 1:
The patent applies local quality by allowing each classifier to have its own threshold index tailored to its specific characteristics and calibration. Instead of a uniform mean-based approach, each classifier's output is adjusted locally based on its threshold index, accounting for distribution differences while maintaining the simplicity of ensemble combination.
Data Source
AI summary
An example system includes a processor to receive training data used to train an ensemble voting classifier. For each classifier in the ensemble voting classifier, the processor can also set a classification score of a positive training item as a threshold. The processor can further adjust a threshold of at least one of the classifiers based on an analysis of a vote contribution of each classifier on the votes on the training data. The threshold of the at least one of the classifiers is adjusted to increase a voting specificity without impacting sensitivity with respect to the training data.