Minority Data Augmentation Using Embedding-Guided GANs
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing oversampling techniques like SMOTE are ineffective in augmenting minor classes in mixed data sets containing both continuous and nominal variables, leading to class imbalance issues in machine learning models.
Innovation Solution
A method and apparatus for data augmentation using a mode separating unit, embedding vector generating unit, auxiliary classifying unit, and classifying unit to generate and refine minor class fake data through generative adversarial networks and metric learning, ensuring data similarity within categories and distance between categories.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Quantity of substance
If SMOTE technique is used for data augmentation, then minor class data is augmented in continuous variable datasets, but it becomes ineffective when nominal variables are mixed with continuous variables
Solution Approach 1:
The patent transforms nominal variables into continuous embedding vectors through neural network encoding, changing the parameter type from discrete categories to continuous numerical representations. This allows the use of continuous variable-based augmentation techniques like GANs on mixed data types that previously could not be handled by SMOTE
Solution Approach 2:
The patent replaces the mathematical interpolation mechanism of SMOTE with a generative adversarial network mechanism. Instead of linear combinations of minority class samples, the system uses a generator network trained to produce realistic minority class samples that pass through a discriminator network, enabling effective augmentation for mixed data types
2Adaptability or versatility
If multiple generating units are used to generate diverse fake data, then data diversity is improved, but system complexity increases due to the need for gating networks and multiple training objectives
Solution Approach 1:
The patent divides the data generation task into multiple specialized generating units, each responsible for generating fake data with specific characteristics or from different latent spaces. The gating network segments the selection process by routing to appropriate generators based on the desired data properties, managing complexity through functional decomposition
Solution Approach 2:
The gating network acts as an intermediary between the latent space and multiple generating units, selecting which generator to use based on the input characteristics. This mediator manages the complexity of multiple generators by providing a unified interface and intelligent routing mechanism
Data Source
AI summary
An apparatus for data augmentation includes a mode separating unit generating minor class fake data from a latent vector, an embedding vector generating unit generating embedding vectors for major class original data, minor class original data, and the minor class fake data through a metric network, an auxiliary classifying unit classifying a class of the embedding vectors from the embedding vector generating unit, and feedbacking the classified result in the mode separating unit, and a classifying unit determining whether the input data is authentic by receiving the embedding vector of the minor class original data and the embedding vector of the minor class fake data from the embedding vector generating unit, and feedbacking the determined result in the mode separating unit.