Incremental ML Model Training for Data Drift Performance Loss
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Machine learning models experience performance degradation over time due to data drift, leading to increased prediction errors, and traditional retraining methods incur high overhead in terms of time and technical resources.
Innovation Solution
Implement incremental training using inference results with feedback data sets to selectively filter records, adjust sample proportions, determine weights, and apply them in a loss function to update the ML model, reducing the need for full retraining.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If traditional retraining approach is used with full training data, then model performance is maintained, but time and technical resources are consumed excessively
Solution Approach 1:
The patent segments the training data into two distinct parts: a historical training data set used for initial model training, and an inference results data set containing corrected predictions from production use. This segmentation enables the model to be retrained incrementally on only the new inference data rather than requiring full retraining on all historical data, thus reducing time and resource consumption while maintaining performance.
Solution Approach 2:
The patent implements a feedback mechanism where inference results from production use are collected, corrected manually or automatically, and fed back into the training process. The corrected inference results form a new training data set that is used to update the model parameters, allowing the model to learn from actual production performance and continuously improve without requiring complete retraining.
2Reliability
If traditional retraining approach is used with full training data, then model accuracy is maintained, but technical resources are expended excessively
Solution Approach 1:
The patent extracts only the necessary new information from production inference results to create a minimized training data set. By taking out only the corrected inference results and excluding redundant historical data from the retraining process, the system achieves model updates with significantly reduced technical resource consumption while maintaining prediction accuracy.
Solution Approach 2:
Instead of performing complete retraining on the full historical data set, the patent applies partial action by training only on the necessary subset of new inference results. This partial retraining approach is sufficient to maintain model accuracy while dramatically reducing the computational resources required compared to full retraining.
3Loss of time
If incremental training with filtered records is used, then training time is reduced, but data processing complexity increases
Solution Approach 1:
The patent performs preliminary actions by pre-processing and correcting inference results before they are used for training. The system prepares the training data in advance by applying correction rules and formatting, which simplifies the subsequent training process and reduces overall data processing complexity despite the incremental approach.
Data Source
AI summary
Methods, systems, and computer-readable storage media directed to incremental training of ML models using incremental data to reduce degradation of performance of the ML models over time. Implementations of the present disclosure reduce overhead, in terms of time and technical resources, expended to retrain ML models using full sets of training data. Further, through incremental training, implementations of the present disclosure evaluate multiple data sets to achieve ML model performance that meets or exceeds performance achieved through training on the entire data sets, while providing reduced training times.