A Review of Time Series Dimensionality Reduction Methods

Dimensionality reduction plays a crucial role in data analysis, helping in the understanding of highly complex, high-dimensional datasets that are often challenging to interpret. Its advantages in simplifying and visualizing intricate data structures are invaluable. The application of dimensionality reduction to time series data can be particularly beneficial, as time series often hold greater complexity compared to other datasets. This complexity arises from the temporal relationships present within individual time series and the interconnected relationships that can exist between multiple time series in a dataset. The goal of this paper is to review more traditional dimensionality reduction approaches to time series analysis but to also offer state of the art alternatives. We explore classical methods such as PCA, MDS, Fourier and Wavelet transforms, and some newer methods like t-SNE, UMAP, ISOMAP and Autoencoders.