Modeling Unevenly Spaced Multivariate Time Series with Mixed Variable Types
Abstract
This research into modeling unevenly spaced multivariate time series with mixed variable types was fruitful and produced 8 submitted/published reports. The most significant work in this research is the unified, flexible and interpretable time series representationframework, Rand-TS, proposed by Gorgulu and Baydogan (2020). This approach handles all the challenges introduced in this research. Rand-TS models density characteristics of the time series data of any type as a mixture of time-varying Gaussian distributions using random decision trees and it embeds density information of each time series into a sparse vector. A simple similarity measure based on this efficient representation is also shown to be successful in classification tasks. Rand-TS is a novel framework for both univariate and multivariate time series (MTS) representation, it can work with time series of various length and missing observations, furthermore, it allows using additional customized features. The experimental results show that Rand-TS provides competitive performance with state-of-the-art similarity-based methods in addition to its flexibility and interpretability. This work was submitted to Data Mining and Knowledge Discovery in April, 2020.
Document Details
- Document Type
- Technical Report
- Publication Date
- Jun 23, 2020
- Accession Number
- AD1108832
Entities
People
- Mustafa Baydogan
Organizations
- Boğaziçi University