Development of high generalization classifiers by controlling margin distributions

2021 Fiscal Year Final Research Report

• Project/Area Number: 19K04441
• Research Category: Grant-in-Aid for Scientific Research (C)
• Allocation Type: Multi-year Fund
• Section: 一般
• Review Section: Basic Section 21040:Control and system engineering-related
• Research Institution: Kobe University
• Principal Investigator: Abe Shigeo 神戸大学, 工学研究科, 名誉教授 (50294195)
• Project Period (FY): 2019-04-01 – 2022-03-31
• Keywords: パターン認識

Outline of Final Research Achievements

To realize a higher generalization ability than that of the conventional support vector machines (SVMs),we developed minimum complexity support vector machines (M SVMs) that fuse conventional SVMs and minimum complexity machines (MCMs) and obtained the following results:
1) Combining standard SVMs (L1 SVMs) and linear programming SVMs (LP SVMs) with MCMs, we developed SL1 SVMs, ML1 SVMs,ML1v SVMs, SLP SVMs, and MLP SVMs. These machines are considered maximizing the minimum margin and minimizing the maximum margin.We developed new training methods for SL1 SVMs, ML1 SVMs,and ML1v SVMs: alternatingly maximizing the minimum margin and minimizing the maximum margin.
2) According to the computer experiments using benchmark problems, ML1v SVMs show the best generalization ability among the developed classifiers and L1 SVMs

Free Research Field

システム制御

Academic Significance and Societal Importance of the Research Achievements

パターン認識アルゴリズムは多くの分野で活用されており、未知のデータに対する高い識別能力、すなわち汎化能力が求められている。このため「多くの分野でSVMより格段に汎化能力の高い識別器が存在しうるか」という学術的な問いに肯定的な答えを求めるべく研究を行った。今回開発したMSVMはその解となりうるものとして学術的意義も高く、また産業界への応用の上でも貢献しうるものである。