中文

WindowMixer:用于时间序列预测的窗口内外建模

机器学习 2024-07-09 v2

摘要

时间序列预测(TSF)在经济预测、天气预报、交通流分析和公共卫生监测等领域至关重要。实际的时间序列数据常包含噪声、异常值和缺失值,这使得精确预测具有挑战性。传统方法仅建模点到点的关系,限制了其捕捉复杂时序模式的能力,并增加了对噪声的敏感性。为此,我们引入基于全MLP框架的WindowMixer模型。WindowMixer利用时间序列的连续特性,从窗口角度考察时序变化。它将时间序列分解为趋势和季节性分量,并分别加以处理。对于趋势分量,采用全连接层进行预测;对于季节性分量,将时间窗口投影为窗口标记(window tokens),并由Intra-Window-Mixer和Inter-Window-Mixer模块进行处理。Intra-Window-Mixer建模窗口内部的关系,而Inter-Window-Mixer建模窗口之间的关系。这种方法能够捕捉数据中复杂的模式和长程依赖。实验表明,WindowMixer在长期和短期预测任务中均优于现有方法。

关键词

引用

@article{arxiv.2406.12921,
  title  = {WindowMixer: Intra-Window and Inter-Window Modeling for Time Series Forecasting},
  author = {Quangao Liu and Ruiqi Li and Maowei Jiang and Wei Yang and Chen Liang and LongLong Pang and Zhuozhang Zou},
  journal= {arXiv preprint arXiv:2406.12921},
  year   = {2024}
}

备注

We have found some errors in the paper, involving inaccurate data, and therefore request to withdraw the manuscript