Correlation and symmetrical uncertainty-based feature selection for multivariate time series classification

Ahmad Saikhu*, Agus Zainal Arifin, Chastine Fatichah

*Corresponding author for this work

Research output: Contribution to journalArticlepeer-review

8 Citations (Scopus)

Abstract

Data relating to everyday phenomena can be recorded in the form of time series. Recently, modeling of time-series has become a topic of interest in data mining. Conducting an analysis of multivariate time series effectively is an essential task for decision-making activities in fields such as meteorology, medicine, and finance. Features selection is a key problem in the analysis of multivariate time series. Rainfall prediction, biomedical classification, pattern recognition, sensor network analysis and so on all have different input features. The problem is that these features have interdependencies and time-delay relationships. Currently, research on the selection of input features of these data still depends on whether they are linear or non-linear. In this paper, we propose a new integration strategy between Pearson Correlation and Symmetrical Uncertainty for relevant feature selection based on linear and non-linear relationships for multivariate time-series classification. We evaluated the goodness of fit of feature subsets using merit value. The meteorological data set was used to test the proposed method. The result showed that the method was able to reduce the number of features by 77.9% features and increase their merit value 2.25 times compared to no input features selection.

Original languageEnglish
Pages (from-to)129-137
Number of pages9
JournalInternational Journal of Intelligent Engineering and Systems
Volume12
Issue number3
DOIs
Publication statusPublished - Jun 2019

Keywords

  • Input selection
  • Irrelevant
  • Linear and non-linear relationship
  • Merit value
  • Multivariate time series
  • Redundant

Fingerprint

Dive into the research topics of 'Correlation and symmetrical uncertainty-based feature selection for multivariate time series classification'. Together they form a unique fingerprint.

Cite this