On the Integration of Large-Scale Time Series Distance Matrices Into Deep Visual Analytic Tools

Time series are essential for modeling a lot of activities such as software behavior, heart rate, and business processes. The analysis of the series data can prevent errors, boost profits, and improve the understanding of behaviors. Among the many techniques available, we can find deep learning techniques and data mining techniques. In data mining, distance matrices between subsequences (similarity matrices, recurrence plots) have already shown their potential in fast large-scale time series behavior analysis. In deep learning, there exist different tools for analyzing the models' embedding space to get insights into the data behavior. DeepVATS is a tool for large time series analysis that allows the visual interaction within the embedding space (latent space) of deep learning models and the original data. The training and analysis of the model may result in a large use of computational resources, resulting in a lack of interactivity. To solve this issue, we integrate distance matrix plots within the tool. The incorporation of these plots with the associated downsampling techniques makes DeepVATS a more efficient and user-friendly tool for a first quick analysis of the data, achieving runtimes reductions of up to 104\documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{mathrsfs} \usepackage{upgreek} \setlength{\oddsidemargin}{-69pt} \begin{document}$$10<^>4$$\end{document} seconds, allowing fast preliminary analysis of datasets of up to 7 M elements. Also, this incorporation allows us to detect trends, extending its capabilities. The new functionality is tested in three use cases: the M-Toy synthetic dataset for anomaly detection, the S3 synthetic dataset for trend detection, and the real-world dataset pulsus paradoxus for anomaly checking.