A deep bidirectional similarity learning model using dimensional reduction for multivariate time series clustering

Abstract

To analyze multivariate time series, research through dimension reduction is being conducted, but flexible dimension reduction cannot be achieved by reflecting the characteristics or types of data. This paper proposed a Deep Bidirectional Similarity Learning model (DBSL) that predicts similarities for multivariate time series clustering. This model is a feature extraction-based on Convolutional Neural Networks (CNN). By setting the filter and pooling size according to the size of the data, the convolution operation for attributes and time series and the pooling process for time series are repeated to perform dimension reduction, and the similarities in the time series are predicted through bidirectional Long Short Term Memory (LSTM). To improve the data noise problem for missing values, which is the biggest problem in time series, a simple moving average was applied to the model. In addition, it deals with the overall type, and the model is not specialized for one data type. The experiment was conducted by classifying the data according to whether it was multivariate or missing, and it was confirmed that the performance of the proposed model was higher than other proposed methods.