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

In long-term time series forecasting (LTSF) tasks, an increasing number of models have acknowledged that discrete time series originate from continuous dynamic systems and have attempted to model their dynamical structures. Recognizing the chaotic nature of real-world data, our model, \textbf{\textit{Attraos}}, incorporates chaos theory into LTSF, perceiving real-world time series as observations from unknown high-dimensional chaotic dynamic systems. Under the concept of attractor invariance, Attraos utilizes non-parametric Phase Space Reconstruction embedding and the proposed multi-scale dynamic memory unit to memorize historical dynamics structure and predicts by a frequency-enhanced local evolution strategy. Detailed theoretical analysis and abundant empirical evidence consistently show that Attraos outperforms various LTSF methods on mainstream LTSF datasets and chaotic datasets with only one-twelfth of the parameters compared to PatchTST.

Paper

References (68)

Scroll for more · 38 remaining

Similar papers

Peer review

Reviewer qpQ46/10 · confidence 2/52024-07-10

Summary

This paper introduces Attraos, a new model for long-term time series forecasting (LTSF) that incorporates chaos theory and views time series data as observations from high-dimensional chaotic dynamic systems. Attraos utilizes attractor invariance, non-parametric Phase Space Reconstruction, and a multi-scale dynamic memory unit to effectively capture historical dynamics and forecast future states with substantially fewer parameters than existing models such as PatchTST. Empirical evidence demonstrates Attraos' superior performance across various LTSF and chaotic datasets, providing a new perspective on the underlying dynamics of time series data.

Strengths

- The paper presents comprehensive experiment results and provide rigorous theoretical results to support their claims. - Leveraging results from chaotic theories for LTSF tasks is a novel and interesting idea. - The method proposed in the paper needs much fewer parameters than previous methods.

Weaknesses

- Most of the descriptions for problem setup (section 2), methods, and theoretical analysis (section 3) are very confusing and hard to follow. There are always notations suddenly coming out without much information. I suspect there are also several typos that cause difficulties in understanding and assessments. See **Questions** for details. - If I understood correctly, the title is a little misleading since seemingly the key insight/perspective comes from the embedding theorem by Takens, a counterpart of Whitney theorem for attractors. Are there any other insights related to chaotic systems? If not, I would suggest concretizing this point in the title, abstract, introduction, and conclusion. - Section A.1 in appendix is entirely copied from Wikipedia, https://en.wikipedia.org/wiki/Takens%27s_theorem , except for some changes on notations. However the notations in this section are not coherent, e.g. $k$ and $N$ in L414. - Similar to the first point, the paper does not provide enough background introduction. If the authors worry containing all the necessary details will exceed the page limit, I would suggest presenting a detailed version of problem setup, backgrounds, analysis and methods in appendix.

Questions

1. L90 Is it a weighted integral $\mu(ds)$ instead of $ds$? For polynomials $\phi$, the integral with Lebesgue measure does not converge. 2. Is $\mu$ same as $\omega$ in L94? 3. Isn’t L94 giving the formula of projection to polynomial subspace and isn’t $K_n$ the kernel? How are they connected to $e^{tA}B$, where $A, B$ only appear in the induced dynamics of $x(t)$? 4. Does $\mathcal{A}_i$ refer that there are multiple attractors in the dynamics? $\mathcal{A}$ are sets, then what does transpose in L158 stand for? 5. L112, 120, if $A$ is a (D,N) tensor, $B$ is a (B,L,N) tensor and $u$ is a $(B,L,D)$ tensor, how is $Ax(t)+Bu(t)$ in (2a) defined? 6. Could the authors give more interpretation on what they intend to do $\Delta$ and $\theta$? I have no idea what the authors mean by ‘$\Delta$ is similar to an attention mechanism’ and how this is related to the previous context. 7. Again, given the shape of $B$ and $\Delta$, it is confusing what $\Delta B$ refers to. 8. How is $H$ in L144 related to previous text?

Rating

6

Confidence

2

Soundness

3

Presentation

1

Contribution

3

Limitations

Yes.

Reviewer zqBV7/10 · confidence 3/52024-07-15

Summary

The paper introduces chaos theory into a long-term time series forcasting (LTSF) model called Attraos (a play on the words "attractor" and "chaos"). They propose a Multi-resolution Dynamic Memory Unit (MDMU) which is inspired by (and looks a lot like) the State Space Models (SSMs) used in the Mamba family of models. However, unlike SSMs, Attraos assumes additional structure on top of time-series signals by utilizing attractor behavior from chaos theory in the phase space. Attraos outperforms many existing architectures (e.g., Mamba, RWKV-TS, PatchTST) on time-series tasks at a fraction of the computational complexity and cost.

Strengths

- S1: **Outperforms existing architectures at a fraction of the cost**. Attraos is able to outperform models like Mamba, RWKV, and Transformers on many different time series tasks at a fraction of the computational cost (in terms of training time and parameter count). - S2: **Theoretically complete**. The paper provide rigorous definitions, theorems, and proofs throughout the main paper and appendix to justify their method. (Note: before this paper I was quite unfamiliar with chaos theory, though I have some experience in SSMs and time-series. I am unable to verify much of the math.) - S3: **Comprehensive empirical results**. The paper compares Attraos to many alternative architectures for time-series data (e.g., RWKV, Mamba, Transformers, MLPs, etc.) and shows that on many tasks Attraos outperforms alternative methods at a fraction of the computational cost. Additionally, the paper includes an ablation study for the different components of Attraos.

Weaknesses

- W1: **Presentation of results could be clearer**. I was often confused by the choice of coloring in the tables. See the following for my suggestions: 1. [Table 1,2,4]. The "Red/blue" colorscheme is unfortunate, because most brains (including mine) associate the color "red" with "poor performance". I suggest using **bold** for best, <u>underline</u> for second best, as is often done in AI papers. 2. [Table 3,5]. For "improvement/decline" in performance, I suggest one of two options: (1) show "better performance" with blue, and "worse performance" with bold and red to emphasize decreased performance from an ablation, or (2) showing only the "deltas" with a "stock ticker arrow" (🔺 or 🔻, colored green for improvement, red for decline). - W2: **Complete architectural description is missing**. Most of the paper focuses on describing how to implement the special components of Attraos, but the complete picture of the architecture is missing. Additionally, there is no section describing the hyperparameters for the training setup for each experiment. See Q1 for specific questions. - W3: **Missing error bars**. Table 1 represents average results of long term forecasting across a swath of model classes, but these averages are not accompanied by reports of standard deviation. Thus, it is difficult to tell whether improved average performance is actually signficant. Improve by including error bars for all numerical results, which may require re-running some experiments under different random seeds. - W4: **Grandiose, non-academic language**. The paper starts with "In the intricate dance of time" [L15] 😂 and makes statements like "we can transcend the limitations of deterministic dynamical systems" [L34]. I suggest reworking these lines to preserve the professionalism of the rest of the work. These weakenesses are admittedly small and easily fixable by the authors during the review process. My overall attitude towards the paper is that it is of high quality and should be accepted. However, a lot of the theory was beyond my ability to evaluate, and I am willing to adjust my score positively/negatively as I become more familiar with this work throughout the review process.

Questions

- Q1: The complete architecture for Attraos is not described. Do you just take Mamba and modify the SSM component? What is the complete set of learnable parameters? Where does the Hopfield Network come into play and how are its memories trained? What are the training hyperparameters for each experiment?

Rating

7

Confidence

3

Soundness

3

Presentation

2

Contribution

3

Limitations

Limitations are adequately described in the conclusion.

Reviewer zUbZ6/10 · confidence 3/52024-07-24

Summary

The paper introduces a novel approach, named Attraos, for Long-term Time Series Forecasting (LTSF) based on treating the observed time series as high dimensional chaotic dynamical system. The model first estimates an embedding of the data through Phase Space Reconstruction (Takens embedding) and then utilizes a memory unit through state space model that represents the dynamical system through polynomials which allows to evolve the dynamics into future steps and forecast the data. The model is evaluated on multiple LTSF datasets and additional experiments such as ablations and robustness under noise are performed.

Strengths

S1. The approach is proposing to forecast data evolution through learning the dynamical system that can generate such time series vs. to predict directly from data. S2. Novel approach for polynomial estimation through a state space model is proposed. S3. Fundamental properties of the approach are rigorously shown. S4. Incorporation of Belloch algorithm for speed up is used. S5. Experiments show that the approach performs better in many cases than existing approaches and ablation experiments are performed.

Weaknesses

W1. There seems to be a detachment between the theorems and properties proved and the proposed system. There seems to be no discussion about how these properties lead to the particular setup in the paper. For example why this particular SSM was used? How the hyperparameters were chosen in light of the propositions? What are the cases that the model is limited and not expected to be effective? Also interpretation of the results is lacking. W2. For non toy experimental results exposition and interpretation of the resulting dynamical system is missing. What is the dynamical system/s that are being obtained for the benchmark and what is the distribution of polynomials?

Questions

See questions in W1 and W2 and also: Q1. The Phase Space Reduction approach seems to be the well known Takens embedding and authors mention Takens thm, however call the method differently. Is there something different that I'm missing in PSR that is not in the classical computation of Takens embedding? Q2. With local evolution I assumed that authors mean single step "autoregressive" forecasting. Is that the case? How would that change for multi-step? Q3. It would be informative if the benefit/limitation/motivation of frequency evol vs. direct evol be more elaborately explained.

Rating

6

Confidence

3

Soundness

3

Presentation

2

Contribution

3

Limitations

See W1-W2 and Q1-Q3. More extended discussion of limitations would contribute to better evaluation of the work vs generalization.

Reviewer zqBV2024-08-08

In this year's review cycle I am unable to see your paper's revisions, but I trust the authors that my suggested changes have been made. Similarly, I thank the authors for answering my questions. After reading the other reviews and responses, I see no reason to decrease my score and keep my vote to accept this paper.

Authorsrebuttal2024-08-10

Thank you for your response! We greatly appreciate your support in recommending our paper! Wishing you all the best!

Reviewer qpQ42024-08-08

Thanks for the clarification. Hope the authors will polish their writing in the final version.

Authorsrebuttal2024-08-10

Thank you for your feedback and validation of our paper! We greatly appreciate your help in improving our scores! Wishing you all the best!

Authorsrebuttal2024-08-12

Kindly Request for Reviewer's Feedback

Dear Reviewer zUbZ, Since the End of author/reviewer discussions is coming in one day, may we know if our response addresses your main concerns? If so, we kindly ask for your reconsideration of the score. Should you have any further advice on the paper and/or our rebuttal, please let us know and we will be more than happy to engage in more discussion and paper improvements. Thank you so much for devoting time to improving our paper!

Reviewer zUbZ2024-08-13

Authors Rebuttal Response

I would like to thank the authors for their concise yet informative rebuttal that have clarified most of my questions. After reading authors rebuttal to my review and other reviewers I remain positive and supportive of the work and novelty of the approach. One issue that I am still wondering about and would like to know if authors are able to further clarify is the detachment of the theory and the approach choices. Could the authors summarize the outcomes of the properties developed in the work and how they led/constrained model choices?

Authorsrebuttal2024-08-13

Thank you for your response and support! We would be delighted to provide you with further clarification on the theory and approach choices. * **Motivation and Significance**: Machine learning has historically been bifurcated into physics-driven approaches (such as PDE numerical solvers, PINN, etc.) and data-driven methods (like large models, various neural networks, neural operators, etc.). While data-driven techniques have excelled in time series modeling (as well as in computer vision, natural language processing, etc.), their black-box nature poses challenges for sustained advancement within the machine learning community. Recent publications in esteemed journals like Nature, Science, PNAS, etc., have showcased a fusion of physics-driven and data-driven methodologies, hinting at the emergence of a prominent research trend involving the integration of physics priors into deep learning models. **Our research marks the inaugural incorporation of PDE dynamical system knowledge into the realm of time series prediction, promising to facilitate interdisciplinary amalgamation and the creation of interpretable deep time series models.** * **Dynamic System**: also known as PDE dynamical systems. Let the domain $S$ be an open subset of $\mathbb{R}^d$ and set an integer $k \geq 1$. Define the system state as $\boldsymbol{x}: S \mapsto \mathbb{R}^m$ where $\boldsymbol{x}=\left(x^1, \ldots, x^m\right)$. Then, an expression of the form:$\mathcal{F}\left(D^k \boldsymbol{x}(s), D^{k-1} \boldsymbol{x}(s), \ldots, D \boldsymbol{x}(s), \boldsymbol{x}(s), s\right)=0$ is called a $k^{\text {th }}$-order system of partial differential equation, where $\mathcal{F}: \mathbb{R}^{m d^k} \times \mathbb{R}^{m d^{k-1}} \times \ldots \times \mathbb{R}^{m d} \times \mathbb{R}^m \times S \mapsto \mathbb{R}^m$ and $s \in S$. real-time series are observed values obtained from the system through a function h. By starting from dynamical systems instead of the original value, it becomes possible to have a better understanding of the fundamental temporal behavior. Moreover, this approach offers the advantages of interpretability and visualization. * **Chaos Theory**: The study of chaos can be considered a branch of dynamical systems, focusing on the fact that both linear and nonlinear dynamical systems tend to exhibit certain fixed shapes in their trajectories, known as attractors. What may appear as irregular behavior in the time domain often reveals stable structures in the dynamical trajectories. The concept of attractors can readily be correlated with current deep-learning pattern recognition technologies. * **Why SSM?**: The reason for choosing SSM to encode dynamical structures is that polynomials are widely recognized as universal approximators in dynamical systems research. The mathematical interpretation of SSM conveniently aligns with polynomial projection, making SSM well-suited for encoding dynamical structures. And we have introduced a novel SSM evolution matrix {-1, -1, -1, ...} to describe finite window approximations. * **Why SSM + Multiple Orthogonal Subspaces**: Building upon SSM, we drew inspiration from chaos-related research (**Theorem 3 Attractor evolution error, line 163**) and proposed an enhanced version of SSM that utilizes distinct orthogonal subspaces to store various attractor structures. This innovation was validated through experimental results. * **Why Frequency Evolution**: In the field of neuroscience research (where EEG and ECG are recognized as chaos datasets), it has been established that attractors are amplified in the frequency domain. Therefore, this paper adopts a frequency-domain evolution strategy. Besides, by discarding high-frequency components, noise effects can be mitigated, reducing computational complexity. Additionally, we explored the effects of Hopfield evolution and KNN clustering evolution, and through experimentation, **we found that these two methods were not as effective as frequency-domain evolution. We provided explanations for these findings in line 257**. * **Follow-up Work**: In our recent research, we have discovered that **incorporating dynamical system priors leads to performance enhancements across all four types of time-series tasks (prediction, classification, interpolation, anomaly detection) and all four neural network architectures (convolutional, attention-based, linear, SSM). These improvements include a tenfold reduction in parameter count, as well as more stable gradients.** This work is set to be released soon—stay tuned for more updates.

Program Chairsdecision2024-09-25

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC