Iterative Machine Learning Model Refinement via Data Block Partitioning
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
In machine learning scenarios, poor quality and low quantity of training data can lead to underfitting or overfitting, and the addition of poor quality data can degrade model accuracy without clear indication until it negatively impacts the model.
Innovation Solution
The method involves partitioning a machine learning model training dataset into multiple training data blocks, generating separate machine learning models for each block, assessing their accuracy, selecting blocks for combination based on accuracy scores, and iteratively combining and relearning data blocks to improve model accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If additional training data is added to a previously trained model, then the model can be updated with new information, but the accuracy of the model can degrade due to poor quality data
Solution Approach 1:
The training dataset is partitioned into multiple blocks, allowing the system to process and evaluate data in manageable segments. This enables selective combination of high-quality data blocks while isolating poor quality data, thus maintaining model accuracy during updates.
Solution Approach 2:
The system generates accuracy scores for models trained on different data blocks and uses this feedback to determine which blocks to combine. This feedback mechanism ensures that only data blocks contributing to model accuracy are integrated, preventing degradation from poor quality data.
2Quantity of substance
If training data blocks are combined to increase data quantity, then the model can leverage more training examples, but poor quality data in the combination can negatively impact model performance
Solution Approach 1:
The system pre-evaluates training data blocks by generating accuracy scores before combining them. This preliminary assessment allows the system to identify and select only high-quality data blocks for combination, ensuring that increased data quantity does not compromise model accuracy.
Solution Approach 2:
The system changes the parameter of data block combination by using accuracy scores as a selection criterion. Instead of combining all available data blocks, it selectively combines blocks that meet accuracy thresholds, thus increasing effective training data quantity while maintaining quality standards.
3Manufacturing precision
If all training data blocks are processed and combined, then comprehensive model training is achieved, but computational resources and time are wasted on poor quality data
Solution Approach 1:
The system extracts and identifies poor quality training data blocks through accuracy score evaluation and excludes them from further processing and combination. This extraction of low-value data prevents waste of computational resources while maintaining comprehensive training on high-quality data.
Solution Approach 2:
Instead of processing all training data blocks equally, the system applies partial action by focusing computational resources only on high-quality data blocks identified through accuracy scoring. This selective processing improves training efficiency without sacrificing model training completeness.
Data Source
AI summary
Technologies are provided for iterative machine learning and relearning. A training dataset can be divided into a plurality of training data blocks which can be used to generate separate machine learning models. The accuracy of the machine learning models can be assessed using a test dataset. Training data blocks which result in models with good accuracy can be combined into larger training data blocks which can then be used to generate new machine learning models. The process of combining training data blocks can be repeated as long as the resulting machine learning model has acceptable accuracy. However, if a model for a combined training data block has a poorer accuracy than the machine learning models for its component training data blocks, then the combined training data block and its machine learning model can be forgotten and its component training data blocks (and their associated machine learning models) can be relearned.