Zipeng365/TSDecompose-Benchmark
收藏资源简介:
TSDecompose基准数据集是一个用于时间序列分解作为独立任务的机制可识别基准测试数据集,源自ICML 2026论文《Time-Series Decomposition as a Standalone Task: A Mechanism-Identifiable Benchmark》。该数据集包含多个数据轨道:1)paper_tables:提供与论文一致的表格数据,包括核心合成基准的原始结果和汇总表格;2)synthetic_full22_extension:扩展的6场景合成数据,使用22种方法进行评估;3)real_proxy22:真实时间序列数据轨道,包含已知周期或机制的数据,用于代理诊断;4)semisynth_transfer:半合成转移数据,将已知机制注入真实背景中进行评估;5)post_rebuttal_second_expansion:反驳后扩展的真实数据证据。核心任务是评估分解方法对时间序列成分(趋势、季节性、残差)的恢复能力,使用五个指标:趋势R2、趋势DTW、季节性R2、季节性谱相关和季节性最大滞后相关。数据集还包括源代码(tsdecomp包)和运行脚本,支持从源文件复现基准测试。数据集适用于时间序列分析、分解方法比较和机制识别研究。
The TSDecompose Benchmark dataset is a mechanism-identifiable benchmark for time-series decomposition as a standalone task, based on the ICML 2026 paper Time-Series Decomposition as a Standalone Task: A Mechanism-Identifiable Benchmark. It includes multiple data tracks: 1) paper_tables: camera-ready manuscript tables with raw and aggregated results from the core synthetic benchmark; 2) synthetic_full22_extension: an extended 6-scenario synthetic benchmark evaluated with a 22-method roster; 3) real_proxy22: a real-data track with public time series featuring known periods or mechanisms, used for proxy diagnostics; 4) semisynth_transfer: a semi-synthetic transfer track injecting known mechanisms into real backgrounds for evaluation; 5) post_rebuttal_second_expansion: a later expansion with real-data evidence from the rebuttal stage. The core task evaluates decomposition methods for component recovery (trend, seasonal, residual) using five metrics: trend R2, trend DTW, seasonal R2, seasonal spectral correlation, and seasonal max-lag correlation. The dataset also includes source code (the tsdecomp package) and scripts to reproduce the benchmark from source. It is designed for time-series analysis, decomposition method comparison, and mechanism identification research.




