A Novel Embedded Discretization-Based Deep Learning Architecture for Multivariate Time Series Classification

Deep learning-based time series classification techniques have significantly improved in recent years. While previous works have mentioned the fundamental importance of temporal discretization, most studies focus on improving model architectures. In this article, several models have been presented that use temporal discretization as a step of preprocessing time series and embed it in the deep neural network. The proposed models consist of two parts: temporal discretization and model training. The first part does the task of discretization and partially the selection of primary features, and the second part does the job of selecting more accurate features and classification. For this purpose, two loss functions have been used: a loss function to evaluate the discretization quality and the other one to evaluate the classification accuracy. The evaluation of the proposed models using 20 benchmarks multivariate time series shows that the proposed methods are more accurate than the state-of-the-art methods.