Learning Data Selection for Estimation Model Accuracy

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing machine learning models struggle to accurately estimate the factors contributing to classification or determination in input data, especially when the data is transformed for processing, leading to reduced determination accuracy.

Innovation Solution

A learning data selection method that calculates distances between input data and transformed data to select appropriate learning target data, ensuring that the estimation model learns from relevant data points in both feature spaces.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If LIME method is used to estimate classification factors in transformed data space, then estimation can be performed on transformed data, but the estimation does not correspond to factors in original input data space

Engineering Contradiction:
Improveclassification factor estimationVSAvoidinput data factor information
Core Design Contradiction:
Measurement precisionVSLoss of information

Solution Approach 1:

The patent introduces an intermediary mapping process that connects the transformed data space back to the input data space. By calculating distances in both spaces and using the transformation relationship, the system mediates between the two feature spaces to identify which input data factors correspond to the transformed data factors, thus preserving the interpretability of original input data while utilizing transformed data for estimation.

Inventive Principle:
Principle #24Intermediary (Mediator)

2Reliability

If data is transformed into dimensions for classifier processing, then classification accuracy is improved, but the relationship between classification factors and original input data becomes unclear

Engineering Contradiction:
Improveclassification accuracyVSAvoidclassification factor identification
Core Design Contradiction:
ReliabilityVSDifficulty of detecting and measuring

Solution Approach 1:

The patent implements a feedback mechanism by calculating distances between input data and transformed data, then using this distance information to identify corresponding factors. The system feeds back the transformation relationship to map classification factors from transformed space back to input data space, making the factors detectable and measurable in terms of original input data while maintaining high classification accuracy.

Inventive Principle:
Principle #23Feedback

Data Source

PatentEP3462386B1Learning data selection program, learning data selection method, and learning data selection device
Publication Date: 2025.06.04 FUJITSU LTD
  • EP3462386B1 patent drawingFigure 1
  • EP3462386B1 patent drawingFigure 2~3
  • EP3462386B1 patent drawingFigure 4

AI summary

To perform selection of learning data of an estimation model that estimates an output factor of a machine learning model that performs classification and determination with respect to transformed data that is transformed from input data, 1) a first input data group relating to first input data is extracted (S12) in correspondence with designation of the first input data included in an input data group input to the machine learning model, 2) a first transformed data group that is input (S14) to the machine learning model and corresponds to the first input data group, and a first output data group of the machine learning model which corresponds to the first transformed data group are respectively acquired, and 3) learning target data of the estimation model is selected (S16) from the first input data group.