Unified formulation and generalization performance of local-feature learning

2019 Fiscal Year Final Research Report

• Project/Area Number: 18K18001
• Research Category: Grant-in-Aid for Early-Career Scientists
• Allocation Type: Multi-year Fund
• Review Section: Basic Section 60010:Theory of informatics-related
• Research Institution: Kyushu University
• Principal Investigator: Suehiro Daiki 九州大学, システム情報科学研究院, 助教 (20786967)
• Project Period (FY): 2018-04-01 – 2020-03-31
• Keywords: 局所パターン / 機械学習 / 汎化性能保証 / shapelet / マルチインスタンス学習

Outline of Final Research Achievements

The main results are as follows:
(1)We proposed a general formulation of the local-feature-based learning problem by using Multiple-Instance Learning framework. (2)We showed the theoretical generalization performance of the local-feature-based hypothesis class. We applied this theory to Shapelet Learning, which is a popular task in the time-series domain, and we gave the first generalization bound of shapelet-based hypothesis class. (3)We proposed an efficient algorithm to solve the learning problem. (4)We demonstrated that our algorithm effectively works in practice.

Free Research Field

統計的学習理論

Academic Significance and Societal Importance of the Research Achievements

本研究は，時系列分類問題におけるShapelet学習のような，局所パターン学習問題と，マルチインスタンス学習問題の関連性を世界で初めて示した．従来各ドメインで独立に発展してきた様々な局所パターン学習問題に対し，統一的な定式化，汎化性能保証，解法を与えたことは，機械学習分野に大きな貢献を与えたと言える．また，本研究の骨子となった「マルチインスタンス学習の枠組みに基づく学習問題の一般化」は，局所パターン学習問題に限らない，様々な学習問題に幅広く展開可能であり，新たな研究分野の開拓に期待できる．