Raster: Representation Learning for Time Series Classification using Scatter Score and Randomized Threshold Exceedance Rate
Alireza Keshavarzian, Shahrokh Valaee · 2023
Time series classification is a fundamental task in many domains such as finance, healthcare, and manufacturing. Traditional machine learning techniques are limited in their ability to deal with the temporal nature of time series data, making representation learning an effective approach. Randomized machine learning is gaining interest as many cases it can outperform deep learning. However, it often requires many features, which can be an issue with high-dimensional training data and low-sample sizes. To address this issue, we propose a novel and efficient approach that utilizes a new and fast metric to evaluate features, called the Scatter Score (SS), and a new temporal-aware down-sampling strategy, called randomized threshold exceedance rate (rTER). Our method achieves significant improvements in classification performance compared to state-of-the-art methods such as ROCKET, miniROCKET, ResNet, and InceptionTime, as demonstrated on 30 different datasets.