TSB-AD benchmark

该数据集名为TSB-AD基准，它包含了多种多样的多变量时间序列数据集，旨在解决影响先前异常检测评估的多种问题，如数据集缺陷、评估措施的偏倚以及不一致的基准测试实践。该基准允许在各种真实世界挑战中进行全面评估，并包括如VUS-PR等评估指标。所有数据集上都使用了一套调整过的超参数，以提高模型的鲁棒性。该数据集的规模涵盖了17个数据集，共180个时间序列，其任务是进行异常检测。

This dataset is named the TSB-AD benchmark. It includes a diverse collection of multivariate time series datasets, designed to resolve multiple problems that plagued prior anomaly detection evaluations, including dataset flaws, biased evaluation metrics, and inconsistent benchmarking practices. This benchmark allows for comprehensive evaluation across various real-world scenarios, and incorporates evaluation metrics such as VUS-PR. A set of tuned hyperparameters was uniformly applied across all datasets to improve model robustness. The benchmark encompasses 17 datasets totaling 180 time series, all targeting the anomaly detection task.

搜集汇总

背景与挑战

背景概述

TSB-AD benchmark是一个包含17个数据集、180个时间序列的多变量时间序列异常检测基准，旨在解决先前评估中的数据集缺陷、评估偏倚和不一致问题，提供全面的评估指标和调整过的超参数以提高模型鲁棒性。