Adversarial Network Synthetic Data Filtering for ML Training
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing systems are limited in their ability to train machine learning models to accurately generate predictions in the event of a data shift between training and production datasets, leading to poor model performance due to discrepancies in data features, target variables, or relationships.
Innovation Solution
The methods and systems use adversarial networks to classify synthetic data, enabling the exclusion of non-representative synthetic data from training datasets, thereby reducing compute resource usage and improving prediction accuracy. Additionally, dynamic selection of training routines based on data shift severity optimizes the training process.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Quantity of substance
If synthetic data is added to training datasets to improve model training, then the amount of training data increases, but compute resource usage increases and training time increases
Solution Approach 1:
The system extracts and removes non-representative synthetic data from the training dataset using adversarial networks. The adversarial network identifies synthetic data samples that do not represent the actual data shift, and these samples are excluded from the training set. This extraction of harmful elements resolves the contradiction by maintaining data volume while eliminating computationally wasteful samples.
Solution Approach 2:
The adversarial network provides feedback mechanisms to evaluate the quality and representativeness of synthetic data. By continuously assessing whether synthetic data samples accurately represent the data shift, the system can iteratively refine the training dataset. This feedback loop ensures that only beneficial synthetic data is retained, optimizing the balance between data volume and training efficiency.
2Reliability
If synthetic data is used to represent data shifts, then prediction accuracy may improve, but the synthetic data may not be representative of the actual data shift
Solution Approach 1:
The adversarial network serves as an intermediary between the synthetic data generator and the training process. It acts as a quality control mechanism that verifies whether synthetic data accurately represents the data shift before the data is used for training. This intermediary layer ensures that only representative synthetic data reaches the training pipeline, resolving the contradiction between improving prediction accuracy and maintaining data shift representation fidelity.
3Adaptability or versatility
If all synthetic data is processed during model training, then comprehensive coverage is achieved, but processing time and computational resources are wasted on non-representative data
Solution Approach 1:
The system performs preliminary filtering of synthetic data using the adversarial network before the main training process begins. By pre-identifying and removing non-representative synthetic data samples, the system avoids wasting training time on problematic data. This preliminary action resolves the contradiction by ensuring comprehensive coverage of representative data while eliminating time-wasting processing of non-representative samples.
Data Source
AI summary
Methods and systems are described herein for facilitating model training related to data shifts. The system may detect, in a production dataset, a data shift from a training dataset used to train a machine learning model. The system may provide the training dataset and the production dataset to an adversarial network to train a first classifier and a second classifier, respectively. The system may provide synthetic data derived from the production dataset to the adversarial network to cause the first classifier and the second classifier to classify the synthetic data. Based on the classifications received from the adversarial network, the system may exclude the synthetic data from an updated training dataset for updating the machine learning model.