Classification Model Training With Feature-Space Minority Augmentation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing classification task models trained with class imbalanced datasets suffer from poor performance due to the dominance of majority class samples, leading to overfitting and reduced accuracy.
Innovation Solution
A method involving a generative adversarial network is used to train a feature generator that augments minority class samples in the feature space, rather than simply duplicating them, and incorporates a data cleaning unit to filter out abnormal vectors, thereby improving the accuracy of the classification task model.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Quantity of substance
If simple duplication of minority class samples is used for upsampling, then the quantity of minority class samples is increased, but the classification task model suffers from overfitting and reduced accuracy
Solution Approach 1:
The patent transforms minority class samples from the original data space to an expanded feature space by learning a mapping function. This parameter transformation allows the same samples to provide more discriminative information without simple duplication, thereby increasing their effective quantity while maintaining model accuracy and preventing overfitting.
Solution Approach 2:
The patent introduces a mapping function as an intermediary between the original minority class samples and the classification task model. This mapping function transforms the samples into an expanded feature space, acting as a mediator that enriches the sample information without requiring direct duplication of the original samples.
2Ease of manufacture
If class imbalanced dataset is used directly for training, then the training process is simple, but the classification task model achieves poor performance
Solution Approach 1:
The patent performs preliminary transformation of minority class samples by learning a mapping to expanded feature space before the actual classification training. This preliminary action prepares the data in advance, ensuring that when the classification model is trained on the class imbalanced dataset, the minority class samples already contain enriched features that improve model performance.
3Reliability
If feature space expansion is used to augment minority class samples, then the accuracy is improved, but the device complexity increases due to generative adversarial network
Solution Approach 1:
The patent extracts and reuses the feature extractor component from the pre-trained classification model and applies it within the generative adversarial network. This extraction allows the same features to serve dual purposes: both in the original classification task and in generating augmented minority class samples, thereby reducing overall system complexity while maintaining accuracy improvements.
Data Source
AI summary
Provided are an artificial intelligence (AI)-based method and apparatus for training a classification task model, a device, and a storage medium, which relate to the field of machine learning (ML) technologies. The method includes: training an initial feature extractor by using a first dataset to obtain a feature extractor, the first dataset being a class imbalanced dataset; constructing a generative adversarial network, the generative adversarial network including the feature extractor and an initial feature generator; training the generative adversarial network by using second class samples to obtain a feature generator; constructing a classification task model, the classification task model including the feature generator and the feature extractor; and training the classification task model by using the first dataset, the feature generator being configured to augment the second class samples in a feature space in a training process of the classification task model.