Multi-domain Time Series Datasets

该数据集包含来自不同领域的时间序列数据集，共计38个，涵盖了人类活动、医疗保健、机械传感器和金融等领域。这些数据集涉及多种任务，包括预测和分类。具体来说，数据集中包含20个预测任务，其预测长度从60到720不等，以及18个分类任务，类别数量从2到52不等。此外，时间序列样本的读出数量（24至1152）和传感器数量（1至963）也各不相同。总的来说，这38个数据集在长度和类别上都具有较大的多样性，任务涵盖了预测和分类两大类。This dataset suite consists of 38 time series datasets from diverse domains, including human activity, healthcare, mechanical sensing, finance, and more. These datasets support multiple task types, primarily forecasting and classification. Specifically, the suite contains 20 forecasting tasks with forecast lengths ranging from 60 to 720, and 18 classification tasks where the number of classes varies from 2 to 52. Additionally, the number of readings per time series sample (ranging from 24 to 1152) and the count of sensors (ranging from 1 to 963) also differ across the datasets. Overall, these 38 datasets exhibit considerable diversity in terms of sequence lengths and class counts, with tasks covering two core categories: forecasting and classification.

搜集汇总

背景与挑战

背景概述

该数据集是UniTS（统一时间序列模型）项目的一部分，包含多领域时间序列数据，用于支持预测、分类、插补和异常检测等多种任务。模型通过统一架构和共享参数处理这些任务，并在38个多领域数据集上表现出色，具备零样本、少样本和提示学习能力。