Sequence Classification Using Third-Order Moments.

Area of Science:

• Machine Learning
• Computational Statistics
• Pattern Recognition

Background:

• Model-based classification of sequence data commonly employs hidden Markov models (HMMs).
• Likelihood-based score functions in HMMs can be computationally intensive, particularly for extended sequences.
• Recent advancements in spectral learning offer new theoretical frameworks for HMM analysis.

Purpose of the Study:

• To develop a more computationally efficient score function for sequence data classification using HMMs.
• To leverage third-order moments inspired by spectral learning for improved classification performance.
• To reduce the computational burden associated with classifying long data sequences.

Main Methods:

• Proposing a novel score function based on third-order moments for HMM classification.
• Utilizing the Kullback-Leibler divergence between theoretical and empirical third-order moments.
• Applying the method to discrete observation sequence data.

Main Results:

• The proposed third-order moment-based method demonstrates lower computational complexity during classification compared to standard likelihood-based techniques.
• Successful classification of both simulated and empirical datasets was achieved.
• The method proved effective in a human activity recognition study.

Conclusions:

• The novel score function based on third-order moments offers a computationally advantageous alternative for sequence data classification with HMMs.
• This approach maintains classification accuracy while significantly reducing computational demands.
• The method shows promise for real-world applications, including human activity recognition.