UCR-Time-Series-Classification-Archive

UCR 时间序列档案 - 于 2002 年推出，已成为时间序列数据挖掘社区的重要资源，至少有 1000 篇发表的论文使用了档案中的至少一个数据集。档案的最初版本有 16 个数据集，但从那时起，它经历了周期性的扩展。上一次扩展发生在 2015 年夏天，当时存档的数据集从 45 个增加到 85 个。本文介绍并将重点关注从 85 个数据集到 128 个数据集的新数据扩展。除了扩展这一宝贵资源之外，本文还为任何希望评估存档新算法的人提供了实用的建议。最后，本文提出了一个新颖但可行的主张：在数百篇论文中显示了对标准基线（1-最近邻分类）的改进，很大一部分可能错误地归因于它们改进的原因。此外，他们可能已经能够通过更简单的修改实现相同的改进，只需要一行代码。

UCR Time Series Archive – launched in 2002, has become a pivotal resource for the time series data mining community, with at least 1,000 published papers leveraging at least one dataset from the archive. The initial version of the archive contained 16 datasets, and it has undergone periodic expansions ever since. The most recent prior expansion occurred in the summer of 2015, when the number of archived datasets rose from 45 to 85. This paper introduces and focuses on the new expansion that increased the archive's dataset count from 85 to 128. Beyond expanding this valuable resource, this paper also provides practical advice for any researcher seeking to evaluate novel algorithms using the archive. Finally, this paper puts forward a novel but feasible assertion: among the hundreds of papers that have reported improvements over the standard baseline (1-nearest neighbor classification), a large proportion may have incorrectly ascribed these performance gains to their purported causal factors. Furthermore, the authors of these studies could have achieved identical improvements with far simpler modifications, requiring only one line of code.