Symbolic Representation for Time Series

This study proposes a novel symbolic representation method for time series data called ASTRIDE. Unlike conventional symbolization techniques, our approach exhibits adaptability in two critical phases: firstly, during the temporal segmentation process, where it dynamically detects change-points in the signals, and secondly, in the sample quantization step, where it leverages quantiles. Additionally, we develop a data-driven edit distance measure for assessing the similarity of our symbolic represen-tations. We demonstrate the performance of our representation compared to standard symbolizations on classification tasks. Our algorithm is evaluated on 86 univariate time series data sets with equal length sourced from the U CR Time Series Classification Archive. An open source GitHub repository is made available to reproduce the experiments in Python.