Title: Motif Discovery using Similarity-Constraints Deep Neural Networks
Authors: Chuitian Rong, Ziliang Chen, Chunbin Lin and Jianming Wang
Abstract: Discovering frequently occurring patterns (or motifs) in time series has many real-life applications in financial data, streaming media data, meteorological data, and sensor data. It is challenging to provide efficient motif discovery algorithms when the time series is big. Existing motif discovery algorithms trying to improve the performance can be classified into two categories: (i) reducing the computation cost but keeping the original time series dimensions; and (ii) applying feature representation models to reduce the dimensions. However, both of them have limitations when scaling to big time series. The performance of the first category algorithms heavily rely on the size of the dimension of the original time series, which performs bad when the time series is big.The second category algorithms cannot guarantee the original similarity properties, which means originally similar patterns may be identified as dissimilar. To address the limitations, we provide an efficient motif discovery algorithm, called FastM, which can reduce dimensions and maintain the similarity properties. FastM extends the deep neural network stacked AutoEncoder by introducing new central loss functions based on labels assigned by clustering algorithms. Comprehensive experimental results on three real-life datasets demonstrate both the high efficiency and accuracy of FastM.
