Time-Delay Ensemble Learning for Robust Model Freshness
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Machine learning models, particularly deep learning models, face issues with model degradation due to data delay, perturbations, and real-time changes, leading to decreased accuracy over time, especially in environments with sudden market or technological changes.
Innovation Solution
An ensemble learning method combining supervised and unsupervised techniques, where time delay is added to training datasets to create an augmented dataset, and multiple models are trained to generate an ensemble machine learning model, reducing sensitivity to data changes and preserving model freshness.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If a machine learning model is trained on a specific dataset to achieve high accuracy, then the model's initial accuracy is improved, but the model's accuracy degrades over time when new data patterns emerge due to model freshness degradation
Solution Approach 1:
The patent implements dynamic model updating by continuously training multiple models on evolving datasets and selecting the best performing model at each time step. This dynamic approach allows the system to adapt to changing data patterns while maintaining high accuracy, resolving the contradiction between initial accuracy and long-term performance stability.
Solution Approach 2:
The system changes the parameter of model training data by incorporating data from multiple time steps (t-1, t-2, etc.) and using augmented datasets that include synthetic samples. This parameter change in training data composition enables the model to maintain accuracy while adapting to new patterns, preventing performance degradation over time.
2Reliability
If a classification model is constantly updated with new data to maintain accuracy, then model freshness is improved, but the training time and computational resources increase
Solution Approach 1:
The patent segments the training process by dividing data into multiple time-step datasets and training separate models for each segment. This segmentation allows parallel training of multiple models on different data segments, reducing the overall training time while maintaining model freshness through selective deployment of the best performing model.
Solution Approach 2:
The system performs preliminary action by pre-training multiple models on historical datasets at different time steps before deployment. This advance preparation allows the system to quickly select and deploy the most appropriate pre-trained model when needed, reducing real-time training requirements and computational overhead.
3Speed
If deep learning models are trained on incomplete or time-sequenced data in real-time environments, then the ability to process real-time data is improved, but model accuracy decreases due to data delay and perturbations
Solution Approach 1:
The patent applies beforehand cushioning by training models on augmented datasets that include synthetic samples and data from multiple time steps. This preparatory cushioning makes the models more robust to data delays and perturbations in real-time environments, maintaining accuracy while enabling real-time processing capability.
Solution Approach 2:
The system uses copying by creating synthetic data samples that replicate real-time data patterns and using these copies for training. This allows the model to learn from comprehensive datasets including edge cases and perturbations without requiring actual real-time data, thereby maintaining accuracy while enabling real-time processing.
Data Source
AI summary
Provided are systems for ensemble learning with machine learning models that include a processor to receive a training dataset of a plurality of data instances, wherein each data instance comprises a time series of data points, add an amount of time delay to one or more data instances to provide an augmented training dataset, select a first plurality of supervised machine learning models, select a second plurality of unsupervised machine learning models, train the first plurality of supervised machine learning models and the second plurality of unsupervised machine learning models based on the augmented training dataset, generate an ensemble machine learning model based on outputs of the supervised machine learning models and unsupervised machine learning models, and generate a runtime output of the ensemble machine learning model based on a runtime input to the ensemble machine learning model. Methods and computer program products are also provided.