A General Architecture for Supervised Classification of Multivariate Time Series

Mohammed Waleed Kadous
Department of Artificial Intelligence
School of Computer Science & Engineering
University of New South Wales
Sydney NSW 2052, Australia
waleed@cse.unsw.edu.au

UNSW-CSE-TR-9806 - September 1998

• Contents
• List of Figures
• List of Tables
• Introduction
  • Goals
• Related work
  • Hidden Markov models
  • Recurrent neural nets
  • Dynamic time warping
  • Recent interest from the AI community
• Distinctions from traditional classification
  • How it is different
    • An indeterminate number of features
    • Many features, not enough data
    • Strong attribute correlation
    • Temporal mismatch
  • New heuristics
    • Temporal Concurrency and Sequences
    • Representing time explicitly
• Formulation of the problem
  • Definition of Terms
    • Channel
    • Stream
    • Frame
    • Stream Set
  • Statement of the problem
  • Assessing success
  • Examples
    • Recognition of signs from a sign language
    • Blues and Reds
• General architecture
  • Overview
  • Training architecture
    • Raw training data
    • Parametrised event primitives
      • Example PEP: A ``straight-line'' approximation for continuous-valued channels
      • Example PEP: ``Delta'' detection for discrete-valued attributes
    • Event extraction
    • Global features
      • Example global attributes
    • Parameter clustering
    • Event attribution
    • Recombination
    • Feature selection
    • Conventional attribute-value learning
    • Classifier
  • Testing architecture
    • Raw test instance
    • Selected event search
    • Selected global feature calculation
    • Recombination
    • Classifier
  • Goal evaluation
    • Generality
    • Classification accuracy
    • Meaningful descriptions
    • Fast learning curve
    • Time
• Example application
  • Domain information
  • Event extraction and global attributes
  • Clustering algorithm
  • Event attribution
  • Feature selection and learning
  • Recognition
  • Results
  • Conclusions
• Conclusions and future work
  • Future work
    • Short term
      • More results, more PEPs, better software
    • Long term
      • Automatic PEP selection and/or generation
      • Making more use of temporal correlations
      • Provable theoretic properties
  • Conclusions
• References
• About this document ...