Similarity Measure of Time Series Based on Siamese and Sequential Neural Networks

Abstract

In order to improve the performance of time series similarity measure, a model combined Siamese and Sequential Neural Network(SSNN) is proposed. The model consists of three parts: siamese neural network, distance measurement and sequential neural network. Time series’ features are extracted through the siamese network, the distance measurement calculates the sequence distance in the feature space, and the sequential network calculates the similarity between two time series according to the result of distance mesurement. The siamese neural network and the sequential neural network are composed of multilayer perceptrons, and the distance measurement calculates the Euclidean distance between the corresponding dimensions of the time series. Experimental results on 22 time series datasets in UCRArchive2018 show that SSNN has better classification performance than several other methods, and has a good ability to measure similarity.