Feature Selection via Dynamic Priority Ranking and Feedback

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing methods for selecting important features in machine learning models are inefficient, often requiring a large number of questions to align with user knowledge, as they may not consistently match on-the-spot knowledge, leading to inconsistent feature selection and increased questioning.

Innovation Solution

An information processing program that prioritizes feature questioning based on estimated importance scores, updating these scores based on user feedback, to efficiently select features consistent with user knowledge by ranking variables and adjusting weights according to user validation.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Extent of automation

If features are selected based on statistical measures (correlation, mutual information, chi-square) or machine learning model impact, then feature selection is automated and systematic, but the number of questions required increases because these methods do not consistently match user's on-the-spot knowledge

Engineering Contradiction:
Improvefeature selection automationVSAvoidnumber of questions
Core Design Contradiction:
Extent of automationVSLoss of time

Solution Approach 1:

The system implements feedback by updating the estimated amount based on user answer results. When users answer questions about feature importance, the system uses these answers to update the estimated importance scores, creating a closed-loop feedback mechanism that progressively refines feature selection to better match user knowledge

Inventive Principle:
Principle #23Feedback

Solution Approach 2:

The feature selection process is made dynamic by allowing the order of priority and estimated amount to change based on user feedback. The system adapts the questioning sequence in real-time, adjusting which features are questioned next based on accumulated user responses rather than following a fixed static order

Inventive Principle:
Principle #15Dynamics

2Productivity

If the maximum possible number of important features are obtained with as few questions as possible, then user satisfaction increases, but the complexity of selecting and prioritizing features increases

Engineering Contradiction:
Improvefeature selection efficiencyVSAvoidfeature prioritization complexity
Core Design Contradiction:
ProductivityVSDevice complexity

Solution Approach 1:

The system performs preliminary action by pre-calculating the order of priority for features before the interactive questioning begins. This preliminary ranking based on statistical measures and model impact provides a starting point that guides the questioning process, reducing the need for exhaustive questioning while maintaining efficiency

Inventive Principle:
Principle #10Preliminary action

Solution Approach 2:

The estimated amount serves as an intermediary that bridges automated feature selection and user knowledge. This intermediate metric combines objective statistical measures with subjective user feedback, allowing the system to balance automation with user satisfaction without requiring complete manual evaluation of all features

Inventive Principle:
Principle #24Intermediary (Mediator)

Data Source

PatentEP4202805A1Information processing program, information processing device, and information processing method
Publication Date: 2023.06.28 FUJITSU LTD
  • EP4202805A1 patent drawingFigure 1
  • EP4202805A1 patent drawingFigure 2~3
  • EP4202805A1 patent drawingFigure 4

AI summary

An information processing program causes a computer(10) to execute a process including, deciding on one or more variables, from among a plurality of variables, to be a target for a question regarding degree of importance, based on order of priority of the variables and estimated amount, the order of priority of the variables being determined based on a plurality of patterns indicating ranking of the plurality of variables, the estimated amount indicating possibility of a match with predetermined condition regarding each of the patterns, (S103) and updating the estimated amount based on an answer result of the question about the decided variable.(S105,S106)