Neural Network Incremental Model Training via Segmented Offline and Online Phases
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing data processing models, particularly those using logistic regression for click-through rate estimation, face challenges in quickly adapting to changing user behaviors and new data due to the time-consuming nature of model training, which prevents timely updates in dynamic online environments.
Innovation Solution
A method involving a neural network where iterative training is performed in stages, using seed models and incremental updates to adapt to new data within specific time windows, allowing for rapid model updates without requiring full retraining, and includes processes for managing failure data to maintain model efficiency and accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If full model training is performed using sufficient training data, then a convergence model that sufficiently fits can be obtained, but model training takes a large amount of time and computational resources
Solution Approach 1:
The patent divides model training into two distinct phases: offline full model training to obtain a converged initial model, and online incremental training to quickly adapt to new data. This segmentation allows the system to leverage the reliability of full training while avoiding its time cost in production environments.
Solution Approach 2:
The patent performs comprehensive model training in advance during an offline phase, obtaining a pre-converged model before deployment. This preliminary action ensures the model is fully trained beforehand, so that during online operation, only quick incremental updates are needed rather than complete retraining.
2Adaptability or versatility
If the model is updated frequently to adapt to changing data environment, then the model can reflect user behavior changes, but full retraining consumes excessive computational resources and time
Solution Approach 1:
The patent implements a dynamic model update strategy where the training approach changes based on the operational phase. In offline mode, full training is performed for comprehensive learning, while in online mode, lightweight incremental training is used for rapid adaptation. This dynamic switching enables frequent updates without excessive resource consumption.
Solution Approach 2:
The patent applies different training qualities to different phases: full-model comprehensive training offline and localized incremental training online. The online phase only trains on new incoming data rather than the entire dataset, providing sufficient adaptability while dramatically reducing computational overhead.
3Measurement precision
If iterative training is performed multiple times to ensure model convergence, then model accuracy is improved, but training time increases significantly
Solution Approach 1:
The patent extracts the intensive iterative training process to an offline phase where time is less critical. The offline training performs multiple iterations to achieve full convergence, while online training performs minimal iterations for quick adaptation. This extraction separates the accuracy-critical iterations from time-sensitive operations.
Data Source
AI summary
A data processing method in a data processing device is provided. First to-be-processed data input into a neural network are obtained. Iterative training is performed on the neural network for a first preset number of times by using first target data in the first to-be-processed data, to obtain a seed model of the neural network. First newly added data generated after an elapse of time corresponding to the first time window is obtained, and the first newly added data and the first to-be-processed data are combined into second to-be-processed data. Iterative training is performed on the seed model for a second preset number of times by using second target data in the second to-be-processed data, to obtain a first incremental model of the neural network. A first preset area overlaps between the second time window and the first time window. The first incremental model online is published.