Date of Award
2017
Document Type
Open Access Thesis
Department
Computer Science and Engineering
Sub-Department
College Of Engineering and Computing
First Advisor
Gabriel Terejanu
Abstract
Time-series analysis is used heavily in modeling and forecasting weather, economics, medical data as well as in various other fields. Change point detection (CPD) means finding abrupt changes in the time-series when the statistical property of a certain part of it starts to differ. CPD has attracted a lot of attention in the artificial intelligence, machine learning and data mining communities. In this thesis, a novel CPD algorithm is introduced for segmenting multivariate time-series data. The proposed algorithm is a general pipeline to process any high dimensional multivariate time-series data using nonlinear non-parametric dynamic system. It consists of manifold learning technique for dimensionality reduction, Gaussian process regression to model the non-linear dynamics of the data and predict the next possible time-step, as well as outlier detection based on Mahalanobis distance to determine the change points. The performance of the new CPD algorithm is assessed on synthetic as well as real-world data for validation. The pipeline is used on federal reserve economic data (FRED) to detect recession. Finally, functional magnetic resonance imaging (fMRI) data of larval zebrafish is used to segment regions of homogeneous brain activity.
Rights
© 2017, Sourav Das
Recommended Citation
Das, S.(2017). Blind Change Point Detection And Regime Segmentation Using Gaussian Process Regression. (Master's thesis). Retrieved from https://scholarcommons.sc.edu/etd/4004