Learning Model Generation for Imbalanced Label Classification

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Conventional machine learning models struggle with overtraining when dealing with imbalanced data sets, particularly in scenarios where one label significantly outnumbers another, leading to biased and inaccurate classification.

Innovation Solution

The approach involves dividing imbalanced training data into subsets, generating multiple learning models based on these subsets, and selecting the most accurate models using evaluation indices, thereby suppressing overtraining and data bias.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Productivity

If machine learning is performed using imbalanced training data where one label significantly outnumbers another, then the learning process completes quickly, but the model becomes overtrained and produces biased classification results

Engineering Contradiction:
Improvelearning speedVSAvoidclassification accuracy
Core Design Contradiction:
ProductivityVSReliability

Solution Approach 1:

The training data set is divided into multiple subsets, where each subset contains a balanced combination of first training data (majority class) and second training data (minority class). This segmentation allows the model to learn from balanced distributions across different subsets rather than being dominated by the majority class in the overall imbalanced data set.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

Each subset is designed with specific local quality characteristics - containing balanced proportions of different labels - while the overall training data remains imbalanced. This enables the model to maintain high classification accuracy for minority classes in each local subset while preserving the global data distribution characteristics.

Inventive Principle:
Principle #3Local quality

2Reliability

If conventional undersampling is used to balance training data, then overtraining is reduced, but data bias is introduced and evaluation index decreases

Engineering Contradiction:
Improvemodel stabilityVSAvoidevaluation index
Core Design Contradiction:
ReliabilityVSMeasurement precision

Solution Approach 1:

Instead of uniformly undersampling the majority class across the entire data set, the data is segmented into multiple subsets with different balanced combinations. This preserves more information from the original data while still achieving balance within each subset, thereby maintaining higher evaluation indices.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The invention changes the parameter of data combination by generating multiple subsets with different balanced ratios and compositions of first and second training data. This allows optimization of both model stability and evaluation index by selecting appropriate subset configurations.

Inventive Principle:
Principle #35Parameter changes

3Quantity of substance

If conventional oversampling with pseudo-modified data is used, then data balance is achieved, but data accuracy is compromised due to uncertain modifications

Engineering Contradiction:
Improvetraining data balanceVSAvoiddata accuracy
Core Design Contradiction:
Quantity of substanceVSMeasurement precision

Solution Approach 1:

The method segments the original training data into multiple subsets without creating pseudo-modified data. Each subset maintains the authenticity of original data while achieving balance through strategic selection and combination of existing first and second training data, thereby preserving data accuracy.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

Instead of creating pseudo-modified copies of minority class data through oversampling, the invention uses copying and recombination of existing balanced combinations of first and second training data across multiple subsets, maintaining data fidelity while achieving balance.

Inventive Principle:
Principle #26Copying

4Device complexity

If a single learning model is generated from imbalanced training data, then the process is simple, but the model exhibits strong data bias and poor generalization

Engineering Contradiction:
Improvemodel generation processVSAvoidmodel generalization
Core Design Contradiction:
Device complexityVSAdaptability or versatility

Solution Approach 1:

The training process is segmented into multiple learning runs, each generating a learning model from a different balanced subset. This segmentation approach maintains relative process simplicity while significantly improving model generalization by exposing the model to various balanced data combinations.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

Multiple learning models generated from different subsets are combined or ensembled to create a final model with improved generalization. This merging of multiple models trained on balanced subsets compensates for the limitations of any single model trained on imbalanced data.

Inventive Principle:
Principle #5Merging (Combining)

Data Source

PatentEP4080422B1Non-transitory computer readable medium, information processing apparatus, and method of generating a learning model
Publication Date: 2026.01.28 YOKOGAWA ELECTRIC CORP
  • EP4080422B1 patent drawingFigure 1
  • EP4080422B1 patent drawingFigure 2
  • EP4080422B1 patent drawingFigure 3

AI summary

A program causes an information processing apparatus to execute operations including determining whether, in a training data set including a plurality of pieces of training data, the count of a first label and the count of a second label are imbalanced, generating, by dividing the training data set, a plurality of subsets each including first training data characterized by the first label and at least a portion of second training data characterized by the second label, the first training data having a count balanced with the count of the second label, generating a plurality of first learning models based on each of the generated subsets, and saving the plurality of first learning models when it is determined that the value of a first evaluation index for the generated plurality of first learning models is higher than the value of a second evaluation index.