Synthetic Data Generation Using Multi-Model Hyperparameter Feedback

The patent divides the synthetic data generation process into multiple independent components: a base distribution model for generating base features, a conditional distribution model for generating conditional features, and a combination model that integrates both. This segmentation allows each component to be optimized independently while working together to produce high-quality synthetic data that closely mimics real data properties.

The patent implements a nested structure where the conditional distribution model operates within the framework of the base distribution model. The base model generates foundational features, and the conditional model builds upon these to generate additional features conditioned on the base features. This nested arrangement enables the system to capture complex dependencies while maintaining computational efficiency.

The patent performs preliminary actions by pre-training the base distribution model and conditional distribution model separately before combining them. The base model is trained to capture fundamental data patterns, and the conditional model is trained to capture relationships between features. This preliminary training enables the combination model to quickly generate accurate synthetic data without requiring extensive computation during the actual generation process.

The patent incorporates feedback mechanisms where the combination model uses predictions from both the base and conditional models to generate synthetic data, and these predictions are continuously refined. The system iteratively updates the models based on the quality of generated data, allowing the models to learn from their own predictions and improve accuracy over time without requiring complete retraining.

The patent designs the base distribution model and conditional distribution model to serve multiple functions. The base model not only generates base features but also provides a framework for understanding overall data patterns. The conditional model not only generates conditional features but also captures relationships between features. This multi-functionality reduces the need for separate specialized models, thereby reducing overall system complexity while maintaining the ability to capture complex probability distributions.

The patent manages system complexity by systematically varying hyperparameters within each model rather than creating entirely separate models for different data characteristics. The base model and conditional model each have their own hyperparameter sets that can be independently optimized. By changing parameters rather than creating new models, the system maintains flexibility in capturing different probability distributions while avoiding the complexity of managing multiple distinct model architectures.