Extension of Incremental Linear Discriminant Analysis for Online Feature Extraction under Unstationary Environments

カテゴリ: 部門大会

論文No: TC15-4

グループ名: 【C】平成24年電気学会電子・情報・システム部門大会講演論文集

発行日: 2012/09/05

タイトル(英語): Extension of Incremental Linear Discriminant Analysis for Online Feature Extraction under Unstationary Environments

著者名: Annie anak Joseph (神戸大学),小澤 誠一(神戸大学)

著者名(英語): Annie anak Joseph (Kobe University),Seiichi Ozawa(Kobe University)

キーワード: Concept Drift|Feature Extraction|Linear Discriminant Analysis|Knowledge Transfer

要約(日本語): One of the recent topics in online learning is the adaptation to “concept drift”, in which an increasing loss of the relevance between the current data to the previous concept representations leads to imposing model changes. On the other hand, online feature extraction for high-dimensional data streams is very important to ensure high-performance and real-time adaptation to dynamic environments. Therefore, we come up with a new approach to an online feature extraction under unstationary environments by extending incremental linear discriminant analysis (ILDA) which can not only detect concept drifts but also perform the knowledge transfer of effective discriminant vectors of the past concepts. The extended ILDA is evaluated for the dataset including sudden change in concepts.

PDFファイルサイズ: 8,471 Kバイト