Synthetic Data Augmentation Using Multi-Model Prediction Sampling
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Real-world data for machine learning training often contains missing or erroneous values, which can lead to errors and decreased performance, and existing methods for dataset augmentation may introduce noise or lose valuable information, especially when using features predicted by other models.
Innovation Solution
A method that uses features predicted by additional machine learning models to generate scores and confidence intervals for unknown features in a training dataset, allowing for the creation of synthetic items and augmentation of the dataset by sampling from these predictions, thereby increasing data size and improving model training.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of manufacture
If data augmentation is performed using expectation values of uncertain ranges, then the processing is simplified, but valuable information is lost
Solution Approach 1:
The patent creates multiple synthetic copies of data items by sampling from uncertainty distributions rather than using single expectation values. Each synthetic item preserves the uncertain range information while maintaining the original data structure, allowing the model to learn from multiple plausible variations without losing the inherent uncertainty information.
Solution Approach 2:
The patent transforms scalar expectation values into multi-dimensional synthetic data samples by introducing a sampling dimension. Instead of processing single expected values, the system generates multiple samples along the uncertainty dimension, effectively adding a new dimension that preserves information about variability and confidence intervals.
2Measurement precision
If multiple machine learning models are used to predict unknown features, then the accuracy and confidence estimation are improved, but the system complexity increases
Solution Approach 1:
The patent combines multiple machine learning models into a unified framework where their predictions are aggregated through sampling from joint uncertainty distributions. Instead of maintaining separate complex systems, the models are merged at the prediction level, sharing a common interface for generating synthetic samples while leveraging the collective strength of multiple models.
Solution Approach 2:
The patent introduces an intermediary sampling mechanism that mediates between multiple model predictions and the final training process. This intermediary layer aggregates predictions from multiple models, samples from their joint uncertainty distributions, and generates synthetic items without requiring direct integration of the models themselves, thus managing complexity.
3Quantity of substance
If synthetic data items are generated to augment the training dataset, then the dataset size is increased, but noise may be introduced
Solution Approach 1:
The patent changes the parameters of synthetic data generation by sampling from uncertainty distributions with varying confidence intervals rather than using fixed transformations. This allows the synthetic data to reflect the actual uncertainty in the predictions, with higher confidence values producing more reliable synthetic samples and lower confidence values producing more conservative augmentations.
Solution Approach 2:
The patent incorporates confidence interval estimates from multiple models as feedback in the synthetic data generation process. Models with higher confidence produce synthetic items that are more likely to be reliable, while lower confidence predictions result in synthetic items that are either excluded or weighted differently during training, creating a feedback loop that maintains data quality.
Data Source
AI summary
A system and a method for training machine learning models, using features predicted by other models, third party models, legacy models, and the like, for training dataset augmentation. Many datasets have many items with unknown, missing, or erroneous values of features. The method comprises using additional machine learning models to predict unknown features, optionally generate a distribution from their inferred score, and use a plurality of scores from the distribution and/or scores from further additional machine learning models, to create replicas for item with missing features, having different estimations of the unknown features. Followingly, train the machine learning model using data items with the known features and the scores for the unknown feature of the item.