Machine Learning Feature Ratio Selection for Normal-Abnormal Classification
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing machine learning models struggle to achieve high accuracy in data processing due to the difficulty in deriving accurate classifiers when using equal numbers of features for normal and abnormal data, despite hyperparameter adjustments.
Innovation Solution
The proposed solution involves selecting a different number of features for normal and abnormal data during training, with the ratio of normal features being between 1.1 to 2 times the abnormal features for one operation and 0.1 to 0.9 times for another operation, followed by feature scaling and mapping to a feature space using kernel functions or neural networks to derive classifiers.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If the same number of features is used for both normal and non-normal product data, then the training data structure is simple and balanced, but the classification accuracy cannot be improved even with hyperparameter adjustments
Solution Approach 1:
The patent applies asymmetry by using different numbers of features for normal and non-normal product data. Specifically, it uses more features for normal products than for non-normal products, creating an asymmetric feature selection strategy that improves classification accuracy while maintaining manageable complexity through systematic selection methods.
Solution Approach 2:
The patent changes the parameter of feature count differently for different classes (normal vs. non-normal products). By adjusting the number of features as a variable parameter based on product type, the system achieves higher classification accuracy without being constrained by a fixed feature count approach.
2Measurement precision
If more features are selected for training, then the model can capture more information, but the training time and computational cost increase
Solution Approach 1:
The patent applies local quality by selecting different numbers of features based on the specific product type. Normal products use more features (higher local quality) while non-normal products use fewer features, optimizing the balance between information capture and training efficiency for each category rather than applying a uniform approach.
Solution Approach 2:
The patent uses partial action by selecting only the necessary number of features for each product type. Rather than using all available features for all products (excessive action), it selectively applies appropriate feature counts, reducing unnecessary computational overhead while maintaining prediction accuracy.
Data Source
AI summary
According to one embodiment, a data processing device includes a processor. The processor acquires first features corresponding to a first classification label and second features corresponding to a second classification label. The processor selects at least a part of the first features from the first features, and at least a part of the second features from the second features. The processor performs a first operation. In the first operation, a first number of the at least a part of the selected first features is not less than 1.1 times and not more than 2 times a second number of the at least a part of the selected second features. The processor generates a first machine learning model based on first training data based on the at least a part of the selected first and second features.