Please use this identifier to cite or link to this item: https://dl.ucsc.cmb.ac.lk/jspui/handle/123456789/4786
Title: History Preservation Approach to Aid in Multivariate Time Series Forecasting
Authors: Mendis, T.K.S
Issue Date: May-2024
Abstract: Abstract Multivariate time series forecasting involves utilizing past and present information to predict future outcomes, enabling informed decision-making. Multivariate time series forecasting holds considerable importance across various domains such as finance, healthcare, and weather forecasting. In multivariate time series forecasting, the accuracy and efficiency of forecasts hinge on how effectively the method captures relationships, correlations, and dependencies from past data. This task is particularly challenging due to the complexity introduced by multiple correlated time series variables. Moreover, significant historical events leave notable impacts on time series data, offering valuable insights that are beneficial if leveraged appropriately. However, existing models in literature make forecasts based on the generalized scenario and lack exploration into investigating and extracting insights from these significant historical events. This study proposes a history preservation technique aimed at identifying and utilizing these events within the forecasting process. This history preservation method uses anomaly detection to identify historically significant events. Then through a data augmentation method, this information is integrated into the original times series dataset. This augmented dataset is then used in forecasting. Experiments were conducted with this method using three machine-learning models and two datasets with two different anomaly detection methods. Results indicated that although there was a decrease in efficiency, accuracy notably improved over the baseline in eight out of a possible twelve instances. Furthermore, five out of the eight improved instances exhibited accuracy improvements exceeding 15% over the baseline. Therefore this approach allows the model to forecast based not only on the generalized scenario but also on significant events that deviate from the norm, leading to greater accuracy in forecasting.
URI: https://dl.ucsc.cmb.ac.lk/jspui/handle/123456789/4786
Appears in Collections:2024

Files in This Item:
File Description SizeFormat 
2019 CS 096.pdf1.75 MBAdobe PDFView/Open


Items in UCSC Digital Library are protected by copyright, with all rights reserved, unless otherwise indicated.