Image-Coded Time Series Classification with MLP-Mixer

Abstract

There are some feature image coding techniques to convert a time series into an image which represents temporal characteristics into spatial information. Convolutional neural network (CNN) based models have been developed for image-coded time series data classification. This paper proposes an MLP-Mixer based model for time series data classification. The proposed model has been compared to a CNN-based model in terms of their image coding and the number of parameters. In the experiments, with fewer parameters, the proposed MLP-Mixer based method has shown comparable performance to the CNN-based model. It also showed that the different combinations of feature image coding could enhance the performance of the classification model.