《计算机应用》唯一官方网站 ›› 2026, Vol. 46 ›› Issue (8): 2457-2466.DOI: 10.11772/j.issn.1001-9081.2025070823

• 人工智能 • 上一篇    下一篇

基于软白化的无负样本自监督学习时序预测框架

李栋, 赵苡积, 丁海燕(), 武浩   

  1. 云南大学 信息学院,昆明 650500
  • 收稿日期:2025-07-23 修回日期:2025-10-13 接受日期:2025-10-14 发布日期:2025-11-05 出版日期:2026-08-10
  • 通讯作者: 丁海燕
  • 作者简介:李栋(1999—),男,上海人,硕士研究生,主要研究方向:时空数据挖掘、网络表征学习
    赵苡积(1993—),男,云南昆明人,讲师,博士,主要研究方向:时空数据挖掘、网络表征学习、服务计算
    丁海燕(1974—),女,云南昆明人,副教授,硕士,主要研究方向:智能数据处理、计算智能
    武浩(1979—),男,河南平顶山人,教授,博士,主要研究方向:Web信息处理、社会媒体分析与服务、服务计算。
  • 基金资助:
    国家自然科学基金资助项目(62362069);云南省基础研究计划项目(202401BF070001-024)

Soft whitening inspired non-contrastive SSL framework for time-series forecasting

Dong LI, Yiji ZHAO, Haiyan DING(), Hao WU   

  1. School of Information Science and Engineering,Yunnan University,Kunming Yunnan 650500,China
  • Received:2025-07-23 Revised:2025-10-13 Accepted:2025-10-14 Online:2025-11-05 Published:2026-08-10
  • Contact: Haiyan DING
  • About author:LI Dong, born in 1999, M. S. candidate. His research interests include spatio-temporal data mining, network representation learning.
    ZHAO Yiji, born in 1993, Ph. D., lecturer. His research interests include spatio-temporal data mining, network representation learning, service computing.
    WU Hao, born in 1979, Ph. D., professor. His research interests include Web information processing, social media analysis and services, service computing.
  • Supported by:
    National Natural Science Foundation of China(62362069);Yunnan Fundamental Research Program(202401BF070001-024)

摘要:

针对现有的自监督学习方法在时序领域存在的依赖负样本对以及数据增强可能破坏时序结构等问题,提出一种基于软白化的无负样本自监督学习时序预测框架(SWTF)。该框架旨在无需负样本即可学习到信息密集且具备预测能力的序列表征。它的核心方法包含两个协同优化目标:一是时序软白化(TSW)损失,即减少表征维度间的冗余并维持表征的方差学习紧凑且信息丰富的表征,从而避免对负样本对和表征不变性约束的依赖;二是去噪重构任务,即通过有针对性的任务设计,引导模型关注并编码具有预测性的时序特征。为了实现上述两个目标的联合优化,构建一个孪生编码器,利用并行多尺度卷积网络高效捕捉从微观的局部周期性到宏观的长程依赖关系。在电力、气象和交通等多个基准数据集上,相较于CoST和PDF(Periodicity Decoupling Framework)等主流时序预测基线,SWTF的平均绝对误差(MAE)降幅约为19.6% (13.4%~26.4%),证实了SWTF在学习鲁棒性和预测性较强的时间序列表征方面的有效性。

关键词: 时序预测, 自监督学习, 软白化, 时序表征学习, 对比学习

Abstract:

To overcome limitations in the existing Self-Supervised Learning (SSL) methods for time-series, such as reliance on negative sample pairs and the potential disruption of temporal structure by data augmentation, a Soft Whitening inspired non-contrastive SSL framework for Time-series Forecasting (SWTF) was proposed. In the framework, series representations with dense information and prediction ability were learned without negative samples. Its core method consists of two collaboratively optimized objectives. The first is a Temporal Soft Whitening (TSW) loss, which means that compact representations with rich information were learned by reducing the redundancy between representation dimensions and maintaining the variance of representations, so as to avoid the reliance on negative sample pairs and representation invariance constraints. The second is a denoising reconstruction task, which means that the model was guided to focus on and encode predictive time-series features. To optimize these two objectives jointly, a Siamese encoder was constructed to use parallel multi-scale convolutional networks to capture both micro-level local periodicities and macro-level long-range dependencies efficiently. Experimental results on several benchmark datasets (electricity, weather, and traffic) show that SWTF achieves a Mean Absolute Error (MAE) reduction of approximately 19.6% (13.4% to 26.4%) compared to mainstream baselines such as CoST and PDF(Periodicity Decoupling Framework), verifying the effectiveness of SWTF in learning robust and highly predictive time-series representations.

Key words: time-series forecasting, Self-Supervised Learning (SSL), soft whitening, time-series representation learning, contrastive learning

中图分类号: