Training Data Generation Device for Feature Imbalance

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing data augmentation methods for machine learning models often lose the original feature imbalance of minority groups, leading to decreased prediction accuracy due to lack of diversity and overtraining.

Innovation Solution

A training data generation device that classifies data into similarity groups based on feature values, selects and converts majority data to maintain the original feature proportions and diversity of minority data, ensuring equivalent data sizes while preserving characteristic imbalances.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Reliability

If training data of minority group is augmented using training data of majority group, then fairness is rectified, but the original feature imbalance of minority group is lost

Engineering Contradiction:
ImprovefairnessVSAvoidfeature imbalance
Core Design Contradiction:
ReliabilityVSLoss of information

Solution Approach 1:

The patent applies local quality by classifying training data into multiple groups based on feature values and selectively converting only specific groups. This allows different parts of the data to be treated differently - majority group data is converted to maintain minority group characteristics while preserving the overall fairness benefit of augmentation.

Inventive Principle:
Principle #3Local quality

Solution Approach 2:

The patent segments the training data into multiple groups based on feature values before performing augmentation. This segmentation enables the system to selectively convert data from specific groups while preserving the feature imbalance characteristics of the minority group, thus resolving the contradiction between fairness and feature imbalance preservation.

Inventive Principle:
Principle #1Segmentation

2Quantity of substance

If training data is augmented to increase data quantity, then model training is improved, but prediction accuracy for minority group decreases due to loss of diversity

Engineering Contradiction:
Improvetraining data quantityVSAvoidprediction accuracy
Core Design Contradiction:
Quantity of substanceVSReliability

Solution Approach 1:

The patent applies local quality by selectively converting only specific groups of data based on their feature values. This ensures that the augmentation process maintains the diversity and characteristic features of minority group data while still increasing the overall data quantity needed for training.

Inventive Principle:
Principle #3Local quality

Solution Approach 2:

The patent performs preliminary classification of training data into multiple groups based on feature values before performing the conversion. This preliminary action allows the system to preserve important feature characteristics during augmentation, preventing loss of diversity that would otherwise reduce prediction accuracy.

Inventive Principle:
Principle #10Preliminary action

3Quantity of substance

If data conversion is performed to generate augmented data, then data quantity is increased, but original feature characteristics may be altered

Engineering Contradiction:
Improveaugmented data quantityVSAvoidfeature characteristics
Core Design Contradiction:
Quantity of substanceVSLoss of information

Solution Approach 1:

The patent applies local quality by converting only specific groups of data identified through feature-based classification. This selective conversion ensures that the augmented data maintains the original feature characteristics of the minority group while still increasing data quantity for training.

Inventive Principle:
Principle #3Local quality

Solution Approach 2:

The patent segments training data into multiple groups based on feature values and applies conversion only to specific segments. This segmentation approach prevents alteration of important feature characteristics during augmentation, as only the appropriate segments are converted while others preserve their original characteristics.

Inventive Principle:
Principle #1Segmentation

Data Source

PatentEP4287075A1Training data generation device and method
Publication Date: 2023.12.06 FUJITSU LTD
  • EP4287075A1 patent drawingFigure 1
  • EP4287075A1 patent drawingFigure 2
  • EP4287075A1 patent drawingFigure 3

AI summary

A training data generation device includes a processor that executes a procedure. The procedure includes classifying, based on a feature value, each of a first plural number of training data having a first attribute and each of a second plural number of training data having a second attribute; based on a comparison of a number of training data classified in a first group against a number of training data classified in a second group from among the first plurality of training data, selecting a third plurality of training data from training data classified in a third group and training data classified in a fourth group from among the second plurality of training data, the third group corresponding to the first group, the fourth group corresponding to the second group; and converting each of the third plurality of training data into a fourth plurality of training data having the first attribute.