Mixed Synthetic Data Generation via Segmented Variational Autoencoder
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for generating synthetic data often fail to effectively augment machine learning training datasets, especially when data is limited and imbalanced, and may not meet privacy requirements, leading to suboptimal model accuracy and implementation challenges with variational autoencoders.
Innovation Solution
A method involving the use of mixed synthetic data generation, where both categorical and continuous variables are encoded, combined, and processed using a machine learning model to reproduce statistical properties of original data, while maintaining privacy by not revealing identifiable information, and utilizing variational autoencoders to improve model accuracy and handle imbalanced datasets.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If existing methods for generating synthetic data are used, then data generation can be performed, but the synthetic data fails to effectively augment machine learning training datasets and may not meet privacy requirements
Solution Approach 1:
The patent segments the synthetic data generation process into distinct components: encoding separate encodings for different data types (categorical, continuous, mixed), processing them through different neural network pathways, and combining them in a decoder. This segmentation allows each component to be optimized for its specific data type while maintaining overall data integrity and statistical properties, thereby improving both model accuracy and data generation effectiveness.
Solution Approach 2:
The patent introduces an intermediary encoding layer that transforms original data into a latent representation space before reconstruction. This intermediary encoding acts as a mediator that preserves statistical properties and privacy while enabling effective data augmentation. The encoded representations are then used to generate synthetic data that maintains the essential characteristics of the original dataset without revealing sensitive information.
2Quantity of substance
If training data is limited, then machine learning model training becomes necessary, but augmenting training data with synthetic data is especially useful when training data is limited
Solution Approach 1:
The patent creates synthetic copies of the original training data through a generative model that learns the statistical properties and data distributions. These synthetic copies augment the limited training data, providing the model with additional examples that maintain the essential characteristics of the original data. This copying approach effectively increases training data quantity while preserving model accuracy by ensuring the synthetic data reflects the true data distribution.
3Ease of manufacture
If variational autoencoders are used for synthetic data generation, then data generation can be performed, but implementation challenges arise
Solution Approach 1:
The patent divides the complex variational autoencoder architecture into separate encoding and decoding pathways, each handling specific data types. The encoder segment processes different data types independently to create type-specific encodings, while the decoder segment reconstructs data from these encodings. This segmentation simplifies implementation by allowing each component to be developed and optimized separately, reducing overall implementation complexity while maintaining architectural sophistication.
Solution Approach 2:
The patent applies different processing approaches and loss functions to different data types within the model. Categorical variables use appropriate encoding and loss functions, continuous variables use different approaches, and mixed variables combine both. This local quality approach allows each part of the model to be optimized for its specific data type, making implementation more straightforward while handling the complexity of diverse data types effectively.
Data Source
AI summary
Methods, systems, and apparatus, including computer programs encoded on a computer storage medium, for generating mixed synthetic data. In one aspect, a method includes obtaining a plurality of mixed input data. At least some of the plurality of mixed input data include one or more categorical variables and one or more continuous variables. The method includes training a machine learning model using the plurality of mixed input data and generating a plurality of mixed synthetic data. The plurality of mixed synthetic data (i) includes one or more categorical variables and one or more continuous variables and (ii) shares statistical properties with the plurality of mixed input data.