Machine Learning Decision Boundary Enhancement
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Machine learning models face inaccuracies in classification due to ambiguous decision boundaries, leading to incorrect classifications of feature vectors near the boundary, which affects the overall accuracy of the models.
Innovation Solution
A method is introduced to enhance decision boundaries by determining a subset of feature sets with classification predictions within a threshold range of the initial decision boundary, using these to generate enhanced training data through a verification model, and further training the machine learning model with both initial and enhanced data to improve classification accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If the machine learning model uses the initial decision boundary for classification, then the classification process is fast, but the classification accuracy is low for feature vectors near the decision boundary
Solution Approach 1:
The patent applies preliminary action by proactively identifying and correcting ambiguous decision boundary regions before they cause misclassifications. The system generates candidate corrections, scores them based on validation performance, and applies the best corrections in advance, preventing accuracy issues rather than reacting to them later.
Solution Approach 2:
The patent replaces the mechanical/brute-force approach of expanding the training dataset with targeted, intelligent corrections to the decision boundary. Instead of mechanically adding more data points, the system uses validation-based scoring and automated correction mechanisms to precisely adjust the boundary where needed, substituting intelligent automation for manual data collection and model retraining.
2Measurement precision
If more training data is collected to improve the decision boundary, then the model accuracy increases, but the time and resources required for training increase
Solution Approach 1:
The patent extracts and focuses computational resources only on the specific regions of the decision boundary that are ambiguous and causing misclassifications. Rather than processing or collecting entire datasets, the system identifies problematic boundary regions, generates targeted corrections for those specific areas, and applies only those necessary changes, significantly reducing training time while maintaining accuracy improvements.
Solution Approach 2:
The patent changes the approach from modifying data parameters (collecting more training data) to modifying model parameters (adjusting the decision boundary directly through corrections). By changing how the model parameters are adjusted—using validation-based scoring to guide precise boundary modifications rather than broad data collection—the system achieves accuracy improvements with minimal training time investment.
3Reliability
If the decision boundary is refined to reduce mis-predictions, then the classification accuracy improves, but the complexity of the verification process increases
Solution Approach 1:
The patent implements feedback mechanisms where the validation model scores candidate decision boundary corrections based on their predicted performance. This feedback loop allows the system to automatically evaluate multiple correction options, select the best ones, and apply them iteratively, improving reliability through systematic validation rather than complex manual verification processes.
Solution Approach 2:
The system applies self-service by automatically generating, scoring, and applying decision boundary corrections without requiring external verification or manual intervention. The validation model autonomously evaluates candidate corrections and the system self-corrects the decision boundary based on validation performance, reducing verification complexity while improving reliability through automated self-improvement.
Data Source
AI summary
Methods, systems, and computer program products for machine learning model decision boundary enhancement are provided. Aspects include determining a decision boundary associated with a first machine learning model, wherein the first machine learning model is trained with an initial training set of features and associated classifications from a verification model, obtaining a plurality of new feature sets, analyzing, by the first machine learning model, the plurality of new feature sets to determine that a subset of feature sets have classification predictions within a threshold range of the decision boundary associated with the first machine learning model, inputting the subset of feature sets into the verification model to determine enhanced training data based on outputs of the verification model, and creating an enhanced machine learning model by further training the first machine learning model with the initial training data and the enhanced training data.