Vectorization Method Based on High Correlation for Multivariate Time Series Hybrid Filter Wrapper Feature Selection

One of the techniques to reduce the multivariate time series dimension is
to transform each Multivariate Time Series (MTS) dataset into a single row or column
called vectorization. This paper contributes to using a new method in forming vectoriza-
tion based on principal component analysis through an observation time analysis factor
of each multivariate time series data without removing any information from the original
data. The vectorization method is called Vectorization for Time of Observation Based on
High Correlation (VecTOR), which is included in the �lter method for feature selection.
The wrapper method selects variables from the vectorization matrix with the Genetic Al-
gorithm { Support Vector Machine algorithm (GASVM). VecTOR-GASVM is compared
to four other methods: VecTOR { Support Vector Machine (VecTOR-SVM), VecTOR-
GABayes, VecTOR Forward-Bayes, and VecTOR Backward-Bayes. The proposed method
has been tested on the CMU and Wafer datasets. Results have shown that the feature
selection of hybrid �lter wrapper VecTOR has fewer features with the highest accuracy
compared to the other four methods. In CMU data, the VecTOR-GASVM method has an
accuracy of 100 per cent with 11 features selected. For the Wafer set of data, VecTOR-
GASVM has an accuracy of 97.98 per cent with 2 features selected.