NeurIPS 2024于2024年12月10号-12月15号在加拿大温哥华举行（Vancouver, Canada）。截止目前，NeurIPS 2024最全论文清单已经重磅出炉了，录取率25.8%。时间序列方向依旧是其中的重要研究领域。

我整理了NeurIPS 2024时间序列（time series data）方向的相关论文，总计61篇，其中正会55篇，D&B Track6篇。

本文列出了包含预测，插补，分类，生成，因果分析，异常检测，LLM以及基础模型、扩散模型的论文标题。需要的同学添加工中号【真AI至上】 回复 NIPS时序 即可全部领取。

Attractor Memory for Long-Term Time Series Forecasting: A Chaos Perspective

文章解析：

本文提出了一种新的模型Attraos，该模型结合了混沌理论，将现实世界的时间序列视为来自未知高维混沌动态系统的观测数据。通过非参数相空间重构嵌入和多尺度动态记忆单元，Attraos能够捕捉历史动态结构，并通过频率增强的局部演化策略进行预测。

实验结果表明，Attraos在主流LTSF数据集和混沌数据集上表现出色，且参数量仅为PatchTST的十二分之一。

创新点：

1.将混沌理论引入长期时间序列预测任务中，利用吸引子不变性的概念建模连续动态系统。

2.提出了一种非参数相空间重构方法，用于恢复时间序列数据中的动态结构。

3.设计了多分辨率动态记忆单元（MDMU），用于捕捉历史采样数据中的结构动态。

4.开发了频率增强的局部演化策略，以在频率域中推导未来状态，提高预测性能。

研究方法：

1.采用非参数相空间重构方法恢复时间序列数据中的动态结构。

2.设计多分辨率动态记忆单元（MDMU）来捕捉历史采样数据中的多尺度动态结构。

3.利用频率增强的局部演化策略，在频率域中推导动态系统组件的未来状态。

4.利用Blelloch扫描算法实现MDMU的高效计算。

研究结论：

1.Attraos在主流LTSF数据集和混沌数据集上表现优于多种LTSF方法，且参数量仅为PatchTST的十二分之一。

2.通过理论分析和实验证据，证明了Attraos的有效性和优越性。

3.Attraos提供了一种新的视角，即利用混沌理论建模和预测时间序列数据。

需要的同学添加工中号【真AI至上】 回复 NIPS时序 即可全部领取。

Frequency Adaptive Normalization For Non-stationary Time Series Forecasting

文章解析：

本文提出了一种新的实例归一化方法——频率自适应归一化（FAN），通过傅里叶变换识别每个输入实例的主要频率成分，并显式建模这些频率成分在输入和输出之间的差异，从而有效处理非平稳时间序列中的动态趋势和季节性模式。FAN方法可以应用于任意预测模型，显著提高了预测性能。

创新点：

1.提出了频率自适应归一化（FAN）方法，能够同时处理动态趋势和季节性模式。

2.利用傅里叶变换识别每个输入实例的主要频率成分，更好地捕捉非平稳因素。

3.显式建模频率成分在输入和输出之间的差异，通过简单的MLP模型进行预测。

4.FAN方法可以应用于任意预测模型，具有广泛的适用性。

研究方法：

1.通过傅里叶变换识别每个输入实例的主要频率成分。

2.显式建模频率成分在输入和输出之间的差异，使用简单的MLP模型进行预测。

3.将FAN方法应用于四个广泛使用的预测模型，评估其在八个基准数据集上的预测性能改进。

4.FAN方法是模型无关的，可以应用于任意预测模型。

研究结论：

1.FAN方法显著提高了预测性能，在MSE指标上平均提升了7.76%到37.90%。

2.与现有的最先进的归一化技术相比，FAN方法表现出优越性。

3.FAN方法不仅能够处理动态趋势，还能有效处理季节性模式，显著提升了非平稳时间序列的预测能力。

NeurIPS 2024 时间序列论文汇总 ↓

预测：

1. Retrieval-Augumented Diffusion Models for Time Series Forecasting

2. Attractor Memory for Long-Term Time Series Forecasting: A Chaos Perspective

3. Attractor Memory for Long-Term Time Series Forecasting: A Chaos Perspective

4. FilterNet: Harnessing Frequency Filters for Time Series Forecasting

5. Frequency Adaptive Normalization For Non-stationary Time Series Forecasting

6. Rethinking the Power of Timestamps for Robust Time Series Forecasting: A Global-Local Fusion Perspective

7. AutoTimes: Autoregressive Time Series Forecasters via Large Language Models

8. DDN: Dual-domain Dynamic Normalization for Non-stationary Time Series Forecasting

9. BackTime: Backdoor Attacks on Multivariate Time Series Forecasting

10. Are Language Models Actually Useful for Time Series Forecasting?

11. Rethinking Fourier Transform for Long-term Time Series Forecasting: A Basis Functions Perspective

12. Introducing Spectral Attention for Long-Range Dependency in Time Series Forecasting

13. Parsimony or Capability? Decomposition Delivers Both in Long-term Time Series Forecasting

14. Structured Matrix Basis for Multivariate Time Series Forecasting with Interpretable Dynamics

15. DeformableTST: Transformer for Time Series Forecasting without Over-reliance on Patching

16. Time-FFM: Towards LM-Empowered Federated Foundation Model for Time Series Forecasting

17. PGN: The RNN's New Successor is Effective for Long-Range Time Series Forecasting

18. SOFTS: Efficient Multivariate Time Series Forecasting with Series-Core Fusion

19. Multivariate Probabilistic Time Series Forecasting with Correlated Errors

20. CycleNet: Enhancing Time Series Forecasting through Modeling Periodic Patterns

21. Analysing Multi-Task Regression via Random Matrix Theory with Application to Time Series Forecasting

22. CondTSF: One-line Plugin of Dataset Condensation for Time Series Forecasting

23. Scaling Law for Time Series Forecasting

24. From News to Forecast: Integrating Event Analysis in LLM-Based Time Series Forecasting with Reflection

25. From Similarity to Superiority: Channel Clustering for Time Series Forecasting

26. TimeXer: Empowering Transformers for Time Series Forecasting with Exogenous Variables

27. ElasTST: Towards Robust Varied-Horizon Forecasting with Elastic Time-Series Transformer

28. Are Self-Attentions Effective for Time Series Forecasting?

29. Tiny Time Mixers (TTMs): Fast Pre-trained Models for Enhanced Zero/Few-Shot Forecasting of Multivariate Time Series

异常检测：

1. SARAD: Spatial Association-Aware Anomaly Detection and Diagnosis for Multivariate Time Series

2. The Elephant in the Room: Towards A Reliable Time-Series Anomaly Detection Benchmark（D&B Track）

分类：

1. Con4m: Context-aware Consistency Learning Framework for Segmented Time Series Classification

2. Medformer: A Multi-Granularity Patching Transformer for Medical Time-Series Classification

3. Abstracted Shapes as Tokens - A Generalizable and Interpretable Model for Time-series Classification

表示学习：

1. "Segment, Shuffle, and Stitch: A Simple Mechanism for Improving Time-Series Representations"

2. Exploiting Representation Curvature for Boundary Detection in Time Series

3. Learning diverse causally emergent representations from time series data

生成：

1. Utilizing Image Transforms and Diffusion Models for Generative Modeling of Short and Long Time Series

2. SDformer: Similarity-driven Discrete Transformer For Time Series Generation

3. FIDE: Frequency-Inflated Conditional Diffusion Model for Extreme-Aware Time Series Generation

4. TSGM: A Flexible Framework for Generative Modeling of Synthetic Time Series（D&B Track）

时序分析：

1. Peri-midFormer: Periodic Pyramid Transformer for Time Series Analysis

2. Shape analysis for time series

3. Large Pre-trained time series models for cross-domain Time series analysis tasks

大语言模型：

1. AutoTimes: Autoregressive Time Series Forecasters via Large Language Models

2. Are Language Models Actually Useful for Time Series Forecasting?

3. From News to Forecast: Integrating Event Analysis in LLM-Based Time Series Forecasting with Reflection

4. Tri-Level Navigator: LLM-Empowered Tri-Level Learning for Time Series OOD Generalization

基础模型：

1. Time-FFM: Towards LM-Empowered Federated Foundation Model for Time Series Forecasting

2. Tiny Time Mixers (TTMs): Fast Pre-trained Models for Enhanced Zero/Few-Shot Forecasting of Multivariate Time Series

3. UNITS: A Unified Multi-Task Time Series Model

4. UniMTS: Unified Pre-training for Motion Time Series

扩散模型：

1. Retrieval-Augumented Diffusion Models for Time Series Forecasting

2. Utilizing Image Transforms and Diffusion Models for Generative Modeling of Short and Long Time Series

3. FIDE: Frequency-Inflated Conditional Diffusion Model for Extreme-Aware Time Series Generation

4. ANT: Adaptive Noise Schedule for Time Series Diffusion Models