Synthetic Training Data Generation for Machine Learning Fairness
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional oversampling techniques for machine learning models, such as FSMOTE, generate synthetic data within a single cluster, which can reduce data variation and lead to overfitting, making it difficult to improve the trade-off between prediction accuracy and fairness, especially when training data is biased towards specific classes or groups.
Innovation Solution
The proposed method involves an information processing apparatus that performs inter-class and inter-group interpolation to generate synthetic training data, balancing cluster sizes by selecting pairs based on neighbor densities and probabilities, thereby preventing overfitting and enhancing the model's generalization performance.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Quantity of substance
If synthetic data is generated within a single cluster using conventional oversampling techniques, then the number of training data for minority classes or groups is increased, but data variation is reduced leading to overfitting
Solution Approach 1:
The patent segments the data generation process into two distinct types of synthetic data: intra-cluster synthetic data (maintaining local characteristics) and inter-cluster synthetic data (introducing global variation). This segmentation allows the system to generate sufficient training data while preserving data diversity by combining both types of synthetic samples.
Solution Approach 2:
The patent applies local quality by generating synthetic data with different levels of variation depending on the source cluster. Intra-cluster synthetic data maintains local characteristics and high fidelity to the source distribution, while inter-cluster synthetic data introduces broader variation. This differentiated approach ensures that synthetic data maintains appropriate local quality while avoiding overfitting.
2Reliability
If synthetic data is generated to balance cluster sizes, then fairness among groups is improved, but prediction accuracy may be reduced due to overfitting
Solution Approach 1:
The patent segments synthetic data generation into intra-cluster and inter-cluster components, where intra-cluster data maintains local fidelity for accuracy while inter-cluster data introduces global diversity for fairness. This segmentation enables the system to achieve both fairness through balanced cluster representation and accuracy through preserved local data characteristics.
Solution Approach 2:
The patent creates a composite synthetic dataset by combining intra-cluster synthetic data and inter-cluster synthetic data. This composite approach integrates the benefits of both generation methods: the local fidelity of intra-cluster synthesis and the global diversity of inter-cluster synthesis, thereby achieving both fairness and prediction accuracy simultaneously.
3Quantity of substance
If conventional oversampling is used to address class imbalance, then minority class representation is improved, but the trade-off between accuracy and fairness cannot be sufficiently improved
Solution Approach 1:
The patent segments the oversampling process into two stages: first generating intra-cluster synthetic data to maintain local characteristics, then generating inter-cluster synthetic data to introduce global variation. This segmented approach enables minority class representation improvement while maintaining generalization performance through diversified synthetic samples.
Solution Approach 2:
The patent extends the synthetic data generation from a single dimension (intra-cluster) to multiple dimensions by incorporating inter-cluster synthesis. This dimensional expansion allows the system to generate synthetic data that not only represents minority classes but also captures broader data distribution patterns, thereby improving generalization performance.
Data Source
AI summary
A non-transitory computer-readable recording medium have stored therein a training data generation program causes a computer to execute a process including, identifying, a first plurality of pieces of training data having label of first value and a first attributes of second value respectively, a second plurality of pieces of training data having the label of the first value and the first attribute of a third values respectively, and a third plurality of pieces of training data having the label of a fourth value and the first attribute of the second value respectively, selecting first training data from among the second or the third plurality of pieces of the training data based on a specific probability, and generating third training data having the label of the first value and the first attribute of the second value based on the first plurality of pieces of training data and the first training data.