Incremental Clustering of Time-series Data based on Self-organizing Incremental Neural Network

• Okada Shogo

  Department of Intelligence Science and Technology, Graduate School of Informatics, Kyoto University

• Nishida Toyoaki

  Department of Intelligence Science and Technology, Graduate School of Informatics, Kyoto University

Bibliographic Information

Other Title

• 自己増殖型ニューラルネットワークを用いた時系列データの追加学習型クラスタリング

Abstract

This paper describes an on-line incremental clustering approach called HMM Based SOINN (HBSOINN) for processing multi variable time-series data such as motion data of gestures. The SOINN (Self-Organizing Incremental Neural Network) is an incremental learning approach that is able to incrementally approximate the distribution of the input data by using efficient numbers of nodes and reporting the number of clusters. We enhanced SOINN by enabling it to cluster time-series data. Hidden Markov Model (HMM) is used to extract features of time-series data and to transform variable-length time-series data to fixed-dimensional data. Experimental results show that HBSOINN outperforms the comparative approach on an artificial data set and 26 kinds of isolated gesture data sets. Even though HBSOINN is an on-line incremental learning approach, it shows the same clustering performance that is evaluated based on the value of Purity and Normalized Mutual Information (NMI) as some state-of-the-art batch clustering approaches.

Journal

• The Brain & Neural Networks

  The Brain & Neural Networks 17 (4), 174-186, 2010

  Japanese Neural Network Society

Details

• CRID: 1390001204467135872
• NII Article ID: 10027705479
• NII Book ID: AA11658570
• DOI: 10.3902/jnns.17.174
• ISSN: 18830455
• ISSN: 1340766X
• Text Lang: ja
• Data Source:
  • JaLC
  • Crossref
  • CiNii Articles
  • KAKEN
• Abstract License Flag: Disallowed