Abstract
Extraction of complex temporal patterns, such as human behaviors, from time series data is a challenging yet important problem. The double articulation analyzer has been previously proposed by Taniguchi et al. to discover a hierarchical structure that leads to complex temporal patterns. It segments time series into hierarchical state subsequences, with the higher level and the lower level analogous to words and phonemes, respectively. The double articulation analyzer approximates the sequences in the lower level by linear functions. However, it is not suitable to model real behaviors since such a linear function is too simple to represent their non-linearity even after the segmentation. Thus, we propose a new method that models the lower segments by fitting autoregressive functions that allows for more complex dynamics, and discovers a hierarchical structure based on these dynamics. To achieve this goal, we propose a method that integrates the beta process - autoregressive hidden Markov model and the double articulation by nested Pitman-Yor language model. Our results showed that the proposed method extracted temporal patterns in both low and high levels from synthesized datasets and a motion capture dataset with smaller errors than those of the double articulation analyzer.
Highlights
In the big data era, we can collect information-rich time series thanks to the advancements in sensing technologies
Beta Process (BP)-autoregressive hidden Markov model (AR-HMM) is applied to time series data, to discover low-level temporal patterns or elemental behaviors (EB), which correspond to the motion primitives in the motion analysis
We introduce the components of our method: BP-AR-HMM and nested Pitman-Yor language model (NPYLM)
Summary
In the big data era, we can collect information-rich time series thanks to the advancements in sensing technologies Such time series data are not segmented and difficult to apply recent machine learning techniques. To segment such data, extraction of temporal patterns in an unsupervised manner is necessary. Many methods have been proposed to extract temporal patterns (Keogh et al, 2004), there exists a problem that the number of existing patterns (and the number of segments) is generally unknown beforehand To solve this issue, non-parametric Bayesian methods are used to determine the number of patterns (Fox et al, 2008b). Non-parametric Bayesian methods based on switching AR models, such as the beta process—autoregressive hidden Markov model (BP-AR-HMM) (Fox et al, 2009, 2014), can be used to identify the temporal patterns without specifying the number of patterns beforehand
Sign-in/Register to view highlights & summary Sign-in/Register to access full text options 
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a callDisclaimer: All third-party content on this website/platform is and will remain the property of their respective owners and is provided on "as is" basis without any warranties, express or implied. Use of third-party content does not indicate any affiliation, sponsorship with or endorsement by them. Any references to third-party content is to identify the corresponding services and shall be considered fair use under The CopyrightLaw.
