Model Relearning Data Selection Using Drift Variation Points
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing machine learning systems face challenges in efficiently relearning due to large amounts of post-drift process data, which can prolong the relearning process and include inappropriate data, leading to decreased model accuracy.
Innovation Solution
A system calculates the inter-distribution distance between target and base data using Sliced Wasserstein Distance (SWD) to identify variation points, determining appropriate training data candidates for relearning by specifying periods with significant changes, and outputs these candidates or meta-information for relearning.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If all post-drift process data is used as training data for relearning, then the learning model can adapt to concept drift, but the relearning process takes a long time
Solution Approach 1:
The patent extracts only the necessary subset of post-drift process data that is appropriate for relearning, rather than using all available data. This is achieved by evaluating data quality and relevance criteria to identify and extract only the useful portion of the data, thereby reducing relearning time while maintaining model accuracy.
Solution Approach 2:
The patent applies partial action by using only a portion of the post-drift process data that meets specific quality criteria. Instead of processing the entire dataset, the system selectively processes only the relevant subset, achieving the necessary model adaptation with reduced computational effort and time.
2Reliability
If all post-drift process data is used as training data, then comprehensive data coverage is achieved, but inappropriate data is included reducing relearning efficiency
Solution Approach 1:
The patent applies local quality by evaluating different portions of the post-drift process data against quality criteria and selecting only those portions that meet the standards. This ensures that high-quality, appropriate data is used for relearning while excluding inappropriate data, thereby improving relearning efficiency without sacrificing model accuracy.
Solution Approach 2:
The system extracts only the appropriate subset of post-drift process data that meets quality criteria, separating useful data from inappropriate data. This extraction process ensures that only relevant, high-quality data is used for training, improving both efficiency and effectiveness of the relearning process.
3Reliability
If a large amount of data is processed for relearning, then model robustness is improved, but processing time increases significantly
Solution Approach 1:
The patent extracts a representative subset of post-drift process data that maintains model robustness while reducing processing time. By applying quality criteria to identify and extract only the necessary data points, the system achieves robust model relearning with significantly reduced computational burden and processing time.
Data Source
AI summary
A system calculates an inter-distribution distance between target data and base data for each of one or more second periods in an entire period between an end time of a first period and a relearning timing of a learning model. For each second period, the target data is data for the second period in the entire data that is data for the entire period. The base data is data for the first period and is data related to creation of the learning model. The system specifies, as a variation point, a time in accordance with the second period during which the calculated inter-distribution distance is greater than or equal to a threshold. The system determines and outputs one or more training data candidates from a part or all of the entire data based on the one or more variation points for relearning of the learning model.