Shadow
收藏资源简介:
Shadow是由蚂蚁集团精心构建的综合性基准数据集,旨在为时间序列预测研究提供贴近现实的评估场景。该数据集融合了开源公共数据集的自然缺失模式与真实工业领域数据,形成了包含多源、不规则采样序列的集合,其规模显著,数据量丰富,能够有效模拟传感器休眠、传输延迟等复杂因素导致的观测不完整性。数据集的创建过程通过整合公开数据与私有数据,并保留其固有的缺失轨迹,以构建具有挑战性的预测任务。该数据集主要应用于时间序列预测领域,特别是针对存在缺失值的场景,旨在评估模型在联合推断未来观测存在性与状态演化方面的能力,为解决现实系统中'信息是否可观测'先于'数值是多少'这一核心问题提供实证基础。
Shadow is a comprehensive benchmark dataset meticulously constructed by Ant Group, designed to provide realistic evaluation scenarios for time series forecasting research. This dataset integrates the natural missing patterns from open public datasets and real industrial data, forming a collection of multi-source, irregularly sampled sequences. It boasts a substantial scale and rich data volume, which can effectively simulate observational incompleteness caused by complex factors such as sensor hibernation and transmission delays. The dataset is developed by integrating public and proprietary data while retaining their inherent missing trajectories, thereby constructing challenging forecasting tasks. This dataset is primarily applied in the field of time series forecasting, particularly for scenarios with missing values, aiming to evaluate the ability of models to jointly infer the existence of future observations and state evolution, and provide an empirical basis for solving the core problem in real-world systems where whether information is observable precedes "what the numerical value is".
数据集概述:Timeflies
核心创新:观察-状态联合建模
Timeflies 突破传统“先插补再预测”和不规则感知范式(假设未来时间戳已知),提出“观察-状态联合范式”:先推断未来观测是否会发生,仅在观测可能有效时预测其值。
方法架构
采用双流管道联合建模观测存在性与演化状态:
- 观测流:从掩码和缺失间隔中捕捉缺失模式。
- 值流:建模时序动态。
关键组件包括:
- 可靠性感知补丁嵌入(Reliability-aware Patch Embedding)
- 观测引导的注意力机制(Observation-guided Value Attention)
- 双预测头(Dual Prediction Head):同时输出未来观测概率与数值预测,适用于真实世界不规则时间序列。
应用场景
适用于未来观测可能不存在的真实世界不规则时间序列预测任务。
相关资源
- 论文:arXiv:2606.13571
- 代码与数据集:基于 Time-Series-Library 和 PatchTST 构建。




