Develop robust classification algorithms for a variety of low-quality data

2021 Fiscal Year Final Research Report

• Project/Area Number: 18K11448
• Research Category: Grant-in-Aid for Scientific Research (C)
• Allocation Type: Multi-year Fund
• Section: 一般
• Review Section: Basic Section 61030:Intelligent informatics-related
• Research Institution: Waseda University
• Principal Investigator: Suko Tota 早稲田大学, 社会科学総合学術院, 准教授 (40409660)
• Project Period (FY): 2018-04-01 – 2022-03-31
• Keywords: パターン認識 / ラベルノイズ / EMアルゴリズム / 漸近解析

Outline of Final Research Achievements

This study deals with classification algorithms that predict labels corresponding to certain features based on accumulated data. In practical applications of classification algorithms, low-quality data including noise is often used.
In this study, we proposed a high-performance classification algorithm that uses a unified model to represent various types of noise. Theoretical performance limits of the proposed algorithm are derived. We analyzed the performance difference between the performance limits and the proposed algorithm.

Free Research Field

機械学習

Academic Significance and Societal Importance of the Research Achievements

現在，画像認識やテキスト分類などの分類アルゴリズムは広く普及しており，一般にも実用されている．しかしながら，実用の場面ではノイズを含む低品質なデータが用いられる事も多く，分類アルゴリズムの本来の性能が発揮できていない場合があり場合によっては十分な分類精度が得られない事がある．本研究の成果を発展させることで，より多くの場面で高性能な分類アルゴリズムが開発できる可能性がある．