Tree-Based Generative Model for Synthetic Data Augmentation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Machine learning models face challenges in training due to limited sample sizes, especially with high-dimensional data structured as trees, leading to overfitting and imbalanced distributions, which complicates generalization and accuracy.
Innovation Solution
A system and method for generating synthetic samples using a tree-based generative model that parameterizes a probability distribution based on conditional probabilities of tree structures, allowing for data augmentation to enhance the training dataset by creating new samples that are contextually relevant and balanced.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Quantity of substance
If data augmentation is performed using traditional methods (e.g., copying and pasting, rotating, flipping), then the training dataset size is increased, but the synthetic samples fail to capture complex associations and hierarchical structures in high-dimensional tree-structured data
Solution Approach 1:
The patent transforms the data augmentation approach by changing the parameterization method from simple geometric transformations to a probabilistic model that captures hierarchical relationships. The generative model uses conditional probabilities parameterized by tree structure to generate synthetic samples that maintain the complex associations and hierarchical structures of the original data, thereby improving both quantity and quality of training samples.
2Reliability
If more training samples are collected to improve model generalization, then the model accuracy improves, but the cost of data collection and annotation increases
Solution Approach 1:
The patent applies the copying principle by generating synthetic samples that replicate the statistical properties and hierarchical structures of real data. Instead of collecting and annotating additional real samples, the system creates artificial copies through the generative model, which captures the essential patterns and relationships in the data without requiring manual annotation efforts.
3Reliability
If the training dataset is expanded with synthetic samples to balance class label distributions, then the model performance on imbalanced classes improves, but the complexity of the data generation process increases
Solution Approach 1:
The patent segments the data generation process into hierarchical levels corresponding to the tree structure. By breaking down the complex generation task into node-level conditional probability calculations, the system can efficiently generate balanced class distributions. Each node in the tree is processed independently with its own conditional probability parameters, making the overall complex process manageable and scalable.
Data Source
AI summary
Techniques facilitating tree-based associative data augmentation are provided. In one example, a computer-implemented method comprises: generating, by a device operatively coupled to a processor, a model probability distribution for one or more measured samples based on conditional probabilities for respective nodes of a tree structure associated with dimensions of the one or more measured samples; and producing, by the device, synthetic samples by drawing from the model probability distribution for at least one of the one or more measured samples.