ICML 2024 时间序列（Time Series）和时空数据（Spatial-Temporal）论文总结【抢先版】

文摘 2024-05-09 07:42 北京

2024 ICML（International Conference on Machine Learning，国际机器学习会议）在2024年7月21日-27日在奥地利维也纳举行。

（好像ICLR24现在正在维也纳开）。

本文总结了ICML 24有关时间序列(Time Series) 和 时空数据(Spatial-temporal) 的相关论文，如有疏漏，欢迎大家补充。

同时我也蹭一下Mamba的热度，放了3篇ICML24接收的Mamba的文章。

时间序列Topic：预测，因果，表示学习，分类，异常检测，插补，生成，不确定性量化，基础模型，大模型

37篇：预测：1-16，表示学习，时序分析：17-22，position paper：23，24（23是大模型，24是无监督异常检测），分类：25，因果：27，28
大语言模型：16,17, 23
基础模型：4, 8, 37, 20
扩散模型：33，34，36

时空数据Topic：时空点过程，时空预测,时空图等

4篇，最后一篇涉及提示微调和大模型相关。

时间序列（Time Series）
An Analysis of Linear Time Series Forecasting Models
Deep Functional Factor Models: Forecasting High-Dimensional Functional Time Series via Bayesian Nonparametric Factorization
Transformers with Loss Shaping Constraints for Long-Term Time Series Forecasting
Unified Training of Universal Time Series Forecasting Transformers
CATS: Enhancing Multivariate Time Series Forecasting by Constructing Auxiliary Time Series as Exogenous Variables
Unlocking the Potential of Transformers in Time Series Forecasting with Sharpness-Aware Minimization and Channel-Wise Attention
SIN: Selective and Interpretable Normalization for Long-Term Time Series Forecasting
A decoder-only foundation model for time-series forecasting
Efficient and Effective Time-Series Forecasting with Spiking Neural Networks
SparseTSF: Modeling Long-term Time Series Forecasting with 1k Parameters
Irregular Multivariate Time Series Forecasting: A Transformable Patching Graph Neural Networks Approach
Learning Optimal Projection for Forecast Reconciliation of Hierarchical Time Series
Time-Series Forecasting for Out-of-Distribution Generalization Using Invariant Learning
Reservoir Computing for Short High-Dimensional Time Series: an Application to SARS-CoV-2 Hospitalization Forecast
Revitalizing Multivariate Time Series Forecasting: Learnable Decomposition with Inter-Series Dependencies and Intra-Series Variations Modeling
S2IP-LLM: Semantic Space Informed Prompt Learning with LLM for Time Series Forecasting
Multi-Patch Prediction: Adapting LLMs for Time Series Representation Learning
TSLANet: Rethinking Transformers for Time Series Representation Learning
MF-CLR: Multi-Frequency Contrastive Learning Representation for Time Series
Timer: Transformers for Time Series at Scale
TimeSiam: A Pre-Training Framework for Siamese Time-Series Modeling
UP2ME: Univariate Pre-training to Multivariate Fine-tuning as a General-purpose Framework for Multivariate Time Series Analysis
Position Paper: What Can Large Language Models Tell Us about Time Series Analysis
Position Paper: Quo Vadis, Unsupervised Time Series Anomaly Detection?
TimeMIL: Advancing Multivariate Time Series Classification via a Time-aware Multiple Instance Learning
Learning Causal Relations from Subsampled Time Series with Two Time-Slices
Discovering Mixtures of Structural Causal Models from Time Series Data
CauDiTS: Causal Disentangled Domain Adaptation of Multivariate Time Series
A Vector Quantization Pretraining Method for EEG Time Series with Random Projection and Phase Alignment
An Empirical Examination of Balancing Strategy for Counterfactual Estimation on Time Series
Bayesian Online Multivariate Time Series Imputation with Functional Decomposition
Conformal prediction for multi-dimensional time-series
Time Weaver: A Conditional Time Series Generation Model
Probabilistic time series modeling with decomposable denoising diffusion model
TimeX++: Learning Time-Series Explanations with Information Bottleneck
Time Series Diffusion in the Frequency Domain
MOMENT: A Family of Open Time-series Foundation Models
时空数据（Spatial-temporal）
Graph-based Time Series Clustering for End-to-End Hierarchical Forecasting
Graph-based Forecasting with Missing Data through Spatiotemporal Downsampling
Beyond Point Prediction: Score Matching-based Pseudolikelihood Estimation of Neural Marked Spatio-Temporal Point Process
A Simple and Universal Prompt-Tuning Framework for Spatio-Temporal Prediction
蹭Mamba热度
Transformers are SSMs: Generalized Models and Efficient Algorithms with Structured State Space Duality
Vision Mamba: Efficient Visual Representation Learning with Bidirectional State Space Model
Can Mamba Learn How To Learn? A Comparative Study on In-Context Learning Tasks

除给ICML官方链接的论文，其余均挂在arXiv或者openreview。

时间序列

1. An Analysis of Linear Time Series Forecasting Models

作者：William Toner · Luke Darlow

关键词：线性模型、时间序列预测、功能等价性、模型比较、闭式解、线性回归、特征归一化、DLinear（AAAI23）、FITS(ICLR24 Spotlight)、RLinear、NLinear（AAAI23）

机构：爱丁堡大学（Edinburgh），华为研究中心（爱丁堡）

链接：https://arxiv.org/abs//2403.14587

解读：AI论文速读 | 线性时间序列预测模型分析

2. Deep Functional Factor Models: Forecasting High-Dimensional Functional Time Series via Bayesian Nonparametric Factorization

作者：Yirui Liu · Xinghao Qiao · Yulong Pei · Liying Wang

机构：伦敦政治经济学院（LSE），埃因霍芬理工大学，利物浦大学（Liverpool）

关键词：预测，贝叶斯非参数模型，可解释性

链接：https://arxiv.org/abs/2305.14543

3. Transformers with Loss Shaping Constraints for Long-Term Time Series Forecasting

作者：Ignacio Hounie · Javier Porras-Valenzuela · Alejandro Ribeiro

机构：宾夕法尼亚大学（UPenn）

关键词：长时预测，约束学习

链接：https://arxiv.org/abs/2402.09373

4. Unified Training of Universal Time Series Forecasting Transformers

作者：Gerald Woo · Chenghao Liu · Akshat Kumar · Caiming Xiong · Silvio Savarese · Doyen Sahoo

机构：Salesforce，新加坡管理大学（SMU）

链接：https://arxiv.org/abs/2402.02592

代码：https://github.com/SalesforceAIResearch/uni2ts

关键词：大规模预训练模型（没有语言，但是够大），时序预测

5. CATS: Enhancing Multivariate Time Series Forecasting by Constructing Auxiliary Time Series as Exogenous Variables

作者：Jiecheng Lu · Xu Han · Sun · Shihao Yang

机构：佐治尼亚理工学院（Gatech），Amazon

链接：https://arxiv.org/abs/2403.01673

关键词：多元时序预测

CATS

6. Unlocking the Potential of Transformers in Time Series Forecasting with Sharpness-Aware Minimization and Channel-Wise Attention

作者：Romain Ilbert · Ambroise Odonnat · Vasilii Feofanov · Aladin Virmaux · Giuseppe Paolo · Themis Palpanas · Ievgen Redko

链接：华为诺亚方舟实验室，LIPADE, Paris Descartes University

关键词：预测，Transformers

链接：https://arxiv.org/abs/2402.10198

代码：https://github.com/romilbert/samformer

7. SIN: Selective and Interpretable Normalization for Long-Term Time Series Forecasting

作者：Lu Han · Han-Jia Ye · De-Chuan Zhan

关键词：长时预测，可解释性

链接：https://icml.cc/virtual/2024/poster/33594

8. A decoder-only foundation model for time-series forecasting

作者：Abhimanyu Das · Weihao Kong · Rajat Sen · Yichen Zhou

关键词：预测，基础模型，decoder-only

链接：https://arxiv.org/abs/2310.10688

这篇比较火爆，三大号机器之心出过报道：
机器之心：2亿参数时序模型替代LLM？谷歌突破性研究被批「犯新手错误」

9. Efficient and Effective Time-Series Forecasting with Spiking Neural Networks

作者：Changze Lv · Yansen Wang · Dongqi Han · Xiaoqing Zheng · Xuanjing Huang · Dongsheng Li

机构：复旦大学，MSRA

关键词：预测，脉冲神经网络

链接：https://arxiv.org/abs/2402.01533

10. SparseTSF: Modeling Long-term Time Series Forecasting with 1k Parameters

作者：Shengsheng Lin · Weiwei Lin · Wentai Wu · Haojun Chen · Junjie Yang

机构：华南理工大学，鹏城实验室，暨南大学

关键词：长时预测

链接：https://arxiv.org/abs/2405.00946

代码：https://github.com/lss-1138/SparseTSF

11. Irregular Multivariate Time Series Forecasting: A Transformable Patching Graph Neural Networks Approach

作者：Weijia Zhang · Chenlong Yin · Hao Liu · Xiaofang Zhou · Hui Xiong

关键词：不规则多元时序预测

链接：https://icml.cc/virtual/2024/poster/33940

12. Learning Optimal Projection for Forecast Reconciliation of Hierarchical Time Series

作者：Asterios Tsiourvas · Wei Sun · Georgia Perakis · Pin-Yu Chen · Yada Zhu

关键词：多层级时间序列预测

链接：https://icml.cc/virtual/2024/poster/34990

13. Time-Series Forecasting for Out-of-Distribution Generalization Using Invariant Learning

作者：haoxin liu · Harshavardhan Kamarthi · Lingkai Kong · Zhiyuan Zhao · Chao Zhang · B. Aditya Prakash

关键词：预测，分布外泛化，不变学习

链接：https://icml.cc/virtual/2024/poster/34011

14. Reservoir Computing for Short High-Dimensional Time Series: an Application to SARS-CoV-2 Hospitalization Forecast

作者：Thomas Ferté · Dutartre Dan · Boris Hejblum · Romain Griffier · Vianney Jouhet · Rodolphe Thiébaut · Pierrick Legrand · Xavier Hinaut

关键词：高维时序，流行病预测

链接：https://icml.cc/virtual/2024/poster/34677

15. Revitalizing Multivariate Time Series Forecasting: Learnable Decomposition with Inter-Series Dependencies and Intra-Series Variations Modeling

作者：Guoqi Yu · Jing Zou · Xiaowei Hu · Angelica I Aviles-Rivero · Jing Qin · Emma, Shujun Wang

机构：香港理工大学（PolyU），电子科技大学，上海AI Lab，剑桥大学

关键词：多元时序预测，时序分解

链接：https://arxiv.org/abs/2402.12694

16. IP-LLM: Semantic Space Informed Prompt Learning with LLM for Time Series Forecasting

作者：Zijie Pan · Yushan Jiang · Sahil Garg · Anderson Schneider · Yuriy Nevmyvaka · Dongjin Song

机构：康涅狄格大学，摩根士丹利

关键词：预测，提示学习，大语言模型

链接：https://arxiv.org/abs/2403.05798

解读：圆圆的算法笔记：时间序列预测+NLP大模型新作：为时序预测自动生成隐式Prompt

S2IPLLM

17. Multi-Patch Prediction: Adapting LLMs for Time Series Representation Learning

作者：Yuxuan Bian · Xuan Ju · Jiangtong Li · Zhijian Xu · Dawei Cheng · Qiang Xu

关键词：表示学习，大语言模型

链接：https://arxiv.org/abs/2402.04852

解读：圆圆的算法笔记: 2篇最新时间序列大模型工作解读

ALLM4TS

18. TSLANet: Rethinking Transformers for Time Series Representation Learning

作者：Emadeldeen Eldele, Mohamed Ragab, Zhenghua Chen, Min Wu, Xiaoli Li

机构：A*Star（新加坡科技研究局）

关键词：表示学习，轻量级模型，自适应频谱块，交互式卷积块，自监督预训练，Transformer，卷积神经网络。

链接：https://arxiv.org/abs/2404.08472

代码：https://github.com/emadeldeen24/TSLANet

TSLANet

19. MF-CLR: Multi-Frequency Contrastive Learning Representation for Time Series

作者：Jufang Duan · wei zheng · Yangzhou Du · Wenfa Wu · Haipeng Jiang · Hongsheng Qi

关键词：对比学习，表示学习

链接：https://icml.cc/virtual/2024/poster/33488

20. Timer: Transformers for Time Series at Scale

作者：Yong Liu · Haoran Zhang · Chenyu Li · Xiangdong Huang · Jianmin Wang · Mingsheng Long

关键词：时间序列分析，基础模型，Transformer，LTSM（大时间序列语言模型），统一时间序列数据集（UTSD）

链接：https://arxiv.org/abs/2402.02368

解读：【重制版】AI论文速读 | 计时器（Timer）：用于大规模时间序列分析的Transformer

Timer

21. TimeSiam: A Pre-Training Framework for Siamese Time-Series Modeling

作者：Jiaxiang Dong, Haixu Wu, Yuxuan Wang, Yunzhong Qiu, Li Zhang, Jianmin Wang, Mingsheng Long

关键词：预训练，时间序列建模

链接：https://arxiv.org/abs/2402.02475

TimeSiam

22. UP2ME: Univariate Pre-training to Multivariate Fine-tuning as a General-purpose Framework for Multivariate Time Series Analysis

作者：Yunhao Zhang · Liu Minghao · Shengyang Zhou · Junchi Yan

关键词：时间序列分析

链接：https://icml.cc/virtual/2024/poster/33686

23. Position Paper: What Can Large Language Models Tell Us about Time Series Analysis

作者：Ming Jin · Yi-Fan Zhang · Wei Chen · Kexin Zhang · Yuxuan Liang · Bin Yang · Jindong Wang · Shirui Pan · Qingsong Wen

关键词：时间序列分析，大语言模型

链接：https://arxiv.org/abs/2402.02713

解读：AI论文速读 | 立场观点:时间序列分析，大模型能告诉我们什么？

LLM和时间序列结合解决现实问题的巨大潜力

24. Position Paper: Quo Vadis, Unsupervised Time Series Anomaly Detection?

作者：M. Saquib Sarfraz · Mei-Yen Chen · Lukas Layer · Kunyu Peng · Marios Koulakis

关键词：异常检测，无监督

链接：https://arxiv.org/abs/2405.02678

25. TimeMIL: Advancing Multivariate Time Series Classification via a Time-aware Multiple Instance Learning

作者：Xiwen Chen · Peijie Qiu · Wenhui Zhu · Huayu Li · Hao Wang · Aristeidis Sotiras · Yalin Wang · Abolfazl Razi

机构：克莱姆森大学，圣路易斯华盛顿大学，亚利桑那州立大学，亚利桑那大学

关键词：多元时间序列分类，多示例学习

链接：https://arxiv.org/abs/2405.03140

代码：https://github.com/xiwenc1/TimeMIL

TimeMIL

26. Learning Causal Relations from Subsampled Time Series with Two Time-Slices

作者：Anpeng Wu · Haoxuan Li · Kun Kuang · zhang keli · Fei Wu

关键词：因果推理，基于拓扑的算法、后代分层拓扑、条件独立准则

链接：https://openreview.net/forum?id=mGmx41FTTy

27. Discovering Mixtures of Structural Causal Models from Time Series Data

作者：Sumanth Varambally · Yian Ma · Rose Yu

机构：加州大学圣地亚哥分校（UCSD）

关键词：结构因果模型（SCM）

链接：https://arxiv.org/abs/2310.06312

28. CauDiTS: Causal Disentangled Domain Adaptation of Multivariate Time Series

作者：Junxin Lu · Shiliang Sun

关键词：因果解纠缠，域适应

链接：https://icml.cc/virtual/2024/poster/33195

29. A Vector Quantization Pretraining Method for EEG Time Series with Random Projection and Phase Alignment

作者：Haokun GUI · Xiucheng Li · Xinyang Chen

关键词：EEG，矢量量化

链接：https://icml.cc/virtual/2024/poster/34865

30. An Empirical Examination of Balancing Strategy for Counterfactual Estimation on Time Series

作者：Qiang Huang · Chuizheng Meng · Defu Cao · Biwei Huang · Yi Chang · Yan Liu

关键词：反事实估计

链接：https://icml.cc/virtual/2024/poster/34183

31. Bayesian Online Multivariate Time Series Imputation with Functional Decomposition

作者：Shikai Fang · Qingsong Wen · Yingtao Luo · Shandian Zhe · Liang Sun

机构：犹他大学（Utah）,松鼠AI，卡耐基梅隆大学（CMU），阿里巴巴达摩院

关键词：插补，高斯过程，不确定性量化

链接：https://arxiv.org/abs/2308.14906

32. Conformal prediction for multi-dimensional time-series

作者：Chen Xu · Hanyang Jiang · Yao Xie

机构：佐治亚理工大学（Gatech）

关键词：共形预测，不确定性量化

链接：https://arxiv.org/abs/2403.03850

代码：https://github.com/hamrel-cxu/MultiDimSPCI

33. Time Weaver: A Conditional Time Series Generation Model

作者：Sai Shankar Narasimhan · Shubhankar Agarwal · Oguzhan Akcin · Sujay Sanghavi · Sandeep Chinchali

机构：德克萨斯大学奥斯汀分校（UTA）

关键词：条件时间序列生成，扩散模型

链接：https://arxiv.org/abs/2403.02682

34. Probabilistic time series modeling with decomposable denoising diffusion model

作者：Tijin Yan · Hengheng Gong · Yongping He · Yufeng Zhan · Yuanqing Xia

关键词：概率时间序列建模，扩散模型

链接：https://icml.cc/virtual/2024/poster/34729

35. TimeX++: Learning Time-Series Explanations with Information Bottleneck

作者：Zichuan Liu · Tianchun Wang · Jimeng Shi · Xu Zheng · Zhuomin Chen · Lei Song · Wenqian Dong · Jayantha Obeysekera · Farhad Shirani · Dongsheng Luo

关键词：可解释性，信息瓶颈

链接：https://icml.cc/virtual/2024/poster/32881

36. Time Series Diffusion in the Frequency Domain

作者：Jonathan Crabbé · Nicolas Huynh · Jan Stanczuk · Mihaela van der Schaar

机构：剑桥大学

关键词：扩散模型，傅里叶分析

链接：https://arxiv.org/abs/2402.05933

代码：https://github.com/JonathanCrabbe/FourierDiffusion

37. MOMENT: A Family of Open Time-series Foundation Models

作者：Mononito Goswami · Arjun Choudhry · Konrad Szafer · Yifu Cai · Shuo Li · Artur Dubrawski

关键词：基础模型

链接：https://arxiv.org/abs/2402.03885

代码：http://anonymous.4open.science/r/BETT-773F/

解读：时序人：MOMENT：CMU发布首个开源的时间序列基础大模型

时空数据

1. Graph-based Time Series Clustering for End-to-End Hierarchical Forecasting

作者：Andrea Cini · Danilo Mandic · Cesare Alippi

机构：提契诺大学，帝国理工学院，米兰理工大学

关键词：时空预测，图结构

链接：https://arxiv.org/abs/2305.19183

2. Graph-based Forecasting with Missing Data through Spatiotemporal Downsampling

作者：Ivan Marisca · Cesare Alippi · Filippo Maria Bianchi

关键词：缺失值下的时空预测，下采样

机构：提契诺大学，米兰理工大学，挪威特罗姆瑟大学，挪威研究中心

链接：https://arxiv.org/abs/2402.10634

3. Beyond Point Prediction: Score Matching-based Pseudolikelihood Estimation of Neural Marked Spatio-Temporal Point Process

作者：Zichong Li · Qunzhi Xu · Zhenghao Xu · Yajun Mei · Tuo Zhao · Hongyuan Zha

机构：佐治尼亚理工学院（Gatech），香港中文大学

关键词：时空点过程（STPP），基于分数，不确定性量化

链接：https://arxiv.org/abs/2310.16310 （标题略有差异，作者一致）

4. A Simple and Universal Prompt-Tuning Framework for Spatio-Temporal Prediction

作者：Zhonghang Li · Lianghao Xia · Yong Xu · Chao Huang

机构：华南理工大学，香港大学

关键词：提示微调，时空预测

链接：https://icml.cc/virtual/2024/poster/32765

Mamba

Mamba进行了自我更新迭代变为了Mamba2接收了（Gu和Dao换了一下作者顺序）

Transformers are SSMs: Generalized Models and Efficient Algorithms with Structured State Space Duality

作者：Tri Dao，Albert Gu

链接：https://icml.cc/virtual/2024/poster/32613

注：现在都是poster，还没有评出来Oral

另外标题带Mamba的还有两篇

Vision Mamba: Efficient Visual Representation Learning with Bidirectional State Space Model

（已经太多号推过这个文章了）

作者：Lianghui Zhu, Bencheng Liao, Qian Zhang, Xinlong Wang, Wenyu Liu, Xinggang Wang

机构：华中科技大学，地平线机器人，北京智源研究院

链接：https://arxiv.org/abs/2401.09417

代码：https://github.com/hustvl/Vim

VisionMamba

Can Mamba Learn How To Learn? A Comparative Study on In-Context Learning Tasks

作者：Jongho Park, Jaeseung Park, Zheyang Xiong, Nayoung Lee, Jaewoong Cho, Samet Oymak, Kangwook Lee, Dimitris Papailiopoulos

机构：蓝洞工作室（做绝地求生即吃鸡那个公司），首尔大学，威斯康辛大学麦迪逊分校，密歇根大学安娜堡分校

链接：https://arxiv.org/abs/2402.04248

代码：https://github.com/krafton-ai/mambaformer-icl

MambaFormer

如果搜索状态空间模型（State space Models&State-space Models）还有7篇，就不赘述了，放两张截图，感兴趣的读者可以自行查阅。

State space Models

State-space Models