To search, Click below search items.


All Published Papers Search Service


A Hybrid Method for Multivariate Time Series Feature Selection


Ani Dijah Rahajoe, Edi Winarko, Suryo Guritno


Vol. 17  No. 3  pp. 103-111


Multivariate time series is usually used in multimedia, finance, medical, gesture and speech recognition. MTS dataset consists of m row and n column. Thus, classification or clustering would have a large size or a high dimensional space. The purpose of feature selection is to reduce the dimensionality without removing any information from the original variable. This paper contributes by means of new feature selection method based on observation times on each of its feature or variable. The proposed filter method uses the resulting factor loadings analysis from principle component analysis (PCA). The idea is to select features based on the time of observation that have most influences on other observation. Only observation times with the highest loading factor value from each principal component are counted. This method is called Feature Selection based Loadings factor (FSBLF) while the GASVM is used as the wrapper method. Fewer observations are included to make the running time faster than when using all observations. This means that only the selected features and observation times are used to perceive grade prediction in testing data. Not all features and observation times are used, and the original (raw/unprocessed) data are used. This method was compared between methods such as without feature selection (classification with SVM), GABayes, FSBLF, FSBLF_GABayes and FSBLF_GASVM. The proposed method has been tested to the CMU dataset and Wafer dataset. Results have shown the selection of a hybrid method FSBLF_GASVM has a few features with higher accuracy and only using the original data until the end of the feature selection process.


Loadings, SVM, genetic algorithm, wrapper, filter, MTS.