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

Summary

This paper proposes a Selective Structured Components-based Neural Network for Long-term Time Series Forecasting

Strengths

1. This paper demonstrates originality by addressing a crucial limitation in existing SOTA methods, maintains high quality through thorough experimentation and clear presentation, offers significant advancements to the field of time series forecasting, and ensures clarity that aids in understanding and reproduction of the work. 2. The motivation of this paper is intuitive and compelling. Given the large model sizes of current SOTA methods like PatchTST, the idea of using a smaller model to achieve comparable or better performance is highly attractive. 3. The experiments are thorough, and the proposed method achieves state-of-the-art performance. The effectiveness of each sub-module is demonstrated through detailed ablation studies. 4. The code is open source and reproducible, with a straightforward and clear usage process.

Weaknesses

1. In the experiment section, it is noted that most papers use the ETT dataset for ablation studies, likely due to its smaller size, which allows for quicker results. However, you chose the ECL and Traffic datasets instead of ETT, which is a more comprehensive and reliable approach. While this choice is commendable, there is no explanation provided for not using the ETT dataset. 2.It would be more informative to report the model size directly in Table 1. Including the model size would provide a clearer comparison with other SOTA methods and highlight the efficiency of your proposed model. 3.Baselines: Some MTSF models based on LLM have been widely applied [1]. If the authors can demonstrate that SSCNN has advantages in both performance and efficiency, this paper will be more convincing. 4.Some extremely lightweight models have also been proven to have satisfactory performance [2] . Compared to these methods, what are the main advantages of SSCNN? [1]One Fits All:Power General Time Series Analysis by Pretrained LM [2]FITS: Modeling Time Series with 10k Parameters

Questions

See weakness

Rating

7

Confidence

4

Soundness

3

Presentation

3

Contribution

3

Limitations

No problem

Summary

This paper identifies data decomposition as a core bottleneck in time series forecasting and proposes a novel model named SSCNN, a decomposition-based model innovatively enhanced with a selection mechanism. SSCNN is specifically designed to adeptly capture complex regularities in data while maintaining a minimal parameter scale. This paper also provides an in-depth comparison between decomposition and patching, examining both capability and parsimony. Comprehensive experiments show the superior performance of SSCNN.

Strengths

Strong Points: 1. The insight of this paper is attractive and compelling. One of the most crucial characteristics of time series is that they can be viewed as composed of components with different natures, e.g., season, trend, and residual. However, this characteristic has been rarely utilized in related works, or it has been implemented in trivial ways. This paper identifies data decomposition as a core bottleneck in time series forecasting and proves its effectiveness. By decomposing complex data into more learnable components, SSCNN achieves state-of-the-art performance with a minimal number of parameters. 2. The writing of this paper is very clear. I can easily follow the author's logic and understand their points. 3. The experimental results are extensive, including overall performance results, ablation studies of each component, hyperparameter experiments, etc., which validate the effectiveness of SSCNN. 4. The code is reproducible and well documented. I have successfully replicated the authors' results. 5. The authors also provide an in-depth comparative analysis and experimental results between patching and decomposition, which help readers understand the advantages of SSCNN’s insights. This paper emphasizes the importance of decomposition in long-term time series forecasting, addressing the analytical gap in feature decomposition for the first time and providing insights into its rationale for capability and parsimony compared to patching.

Weaknesses

I have some minor questions and suggestions. If the author addresses the following points, I will increase my score. Weak Points: Experimental Setting: Most works using the Time-Series-Library repository predict up to 720 steps, yet your results do not include this prediction horizon. It would be beneficial to explain why 720-step predictions were not included. Figures: I suggest the authors add more explanatory information to Figure 1 to help readers grasp the main architecture of SSCNN from the figure and its caption alone. Moreover, some font styles (italic) in Figure 1 seem different from the character styles in the main text. I recommend unifying the styles. Minor Issues: The operator $\lfloor \cdot \rfloor$ is used in the paper but not explained. In Figure 3(a), if I understand correctly, “HDformer” should be replaced by “SSCNN.” Figures: The text size of the legends in the figures is too small, making them difficult to read. Adjusting the text size to be consistent with the main text would enhance the readability of the figures and improve the overall presentation quality of the paper.

Questions

See Weakness.

Rating

6

Confidence

4

Soundness

3

Presentation

3

Contribution

3

Limitations

Yes

Summary

This paper addresses long-term time series forecasting and critiques the reliance on complex models with extensive parameters. It proposes a decomposition method specifically designed for time series dynamics, achieving better forecasting performance across various datasets. Remarkably, the new model uses over 99% fewer parameters than other methods, highlighting the efficiency of domain-specific approaches. This research calls for a move away from complexity in LTSF, showcasing the effectiveness of focused decomposition techniques rather than relying on large-scale models.

Strengths

1. The paper is praiseworthy for its intuitive approach. It tackles a significant problem by proposing a method that matches or surpasses current state-of-the-art models like PatchTST while using a smaller model footprint. The experimental results strongly validate this approach. 2. The model consistently performs well under various experimental conditions, including different input window sizes and hyperparameter settings. Statistical tests demonstrate the reliability of the results across multiple initializations, strengthening the study's credibility. 3. The authors provide a thorough comparison between decomposition and patching in terms of effectiveness and simplicity, demonstrating the superior benefits of decomposition over patching.

Weaknesses

1. The clarity of the methodology could be improved with further elaboration. 2. The evaluation could be strengthened by including comparisons with LLM-based models, such as: [1] Jin, Ming, et al. "Time-LLM: Time Series Forecasting by Reprogramming Large Language Models." The Twelfth International Conference on Learning Representations. [2] Bian, Yuxuan, et al. "Multi-patch prediction: Adapting llms for time series representation learning." arXiv preprint arXiv:2402.04852 (2024).

Questions

1. The methodology would benefit from additional explanation regarding the structure and rationale of the Polynomial Regression layer depicted in Figure 1. If this layer represents a standard approach, including references would enhance clarity. 2. Clarifying the decision to exclude an attention mechanism from the long-term component, despite its presence in other components like seasonal, short-term, and spatial, would strengthen the methodological coherence and aid reader comprehension. 3. Figure 1 requires clarification on several elements. Specifically, the purpose of the 4x4 blocks and addressing inconsistent text formatting (e.g., $\mathcal{E}$) compared to the main text would improve comprehensibility.

Rating

7

Confidence

4

Soundness

3

Presentation

3

Contribution

3

Limitations

Whether this model can be applied to other types of time series data, e.g. trajectory.

Summary

This study unveils a groundbreaking approach to time series forecasting, notable for its minimal parameter count. It stands as the first model to consistently outperform state-of-the-art (SOTA) techniques while remaining compact. Unlike prevalent methods such as PatchTST and iTransformer, which are powerful but cumbersome, and emerging methods like TimeMixer and SCNN, which are lightweight yet inadequate for complex tasks, this model achieves superior performance without the associated heft.

Strengths

1. The model consistently delivers superior accuracy compared to state-of-the-art (SOTA) methods while maintaining a minimal model size. This accomplishment distinguishes it from other methods. 2. The framework unifies the ability to capture various patterns in time series data, offering a streamlined and enhanced alternative to existing models built with MLPs or Transformers. 3. The authors conduct extensive experiments, showcasing the model's strong performance compared to selected SOTA models, which are sufficiently representative of the latest advancements in the field.

Weaknesses

1. There is a gap between the introduction and Section 3 regarding the decomposition of the time series into four components. The authors do not explain why these four components are sufficient. For longer sequences, is there a need for more components? Are there references that support this approach? This discussion should be included at the beginning of Section 3. 2. Manually disabling the spatial component for certain datasets appears suboptimal. It would be more effective if the algorithm could automatically determine whether including the spatial component is beneficial for each dataset. 3. The paper's formatting needs improvements. It seems the authors may have additional content to include. Although the figures in the methodology section are clear and informative, resizing and rearranging them could provide more space for adding valuable content to the main text.

Questions

If the authors can address the weak points, I would reconsider the score.

Rating

7

Confidence

4

Soundness

3

Presentation

4

Contribution

3

Limitations

The authors have raised the limitation of the model concerning computational efficiency, along with potential solutions.

Summary

Title: Parsimony or Capability? Decomposition delivers both in long term time series forecasting. Long term time series forecasting has been an important research problem which applies to different problem domains. This paper proposes a decomposition method which shows significant performance on the benchmarks with less parameters. This method been evaluated extensively on the various datasets and been competitive to existing models. With such approach models can be enhanced to adapt domain characteristics more effectively in various time series applications.

Strengths

1. SSCNN reduces the parameter count substantially compared to traditional models, holding onto less than 1% of the parameters while still performing well across different datasets. 2. The model captures complex data patterns effectively using fewer parameters, utilizing a structured component-based approach with a selection mechanism to improve prediction accuracy. 3. SSCNN excels in time series forecasting, managing diverse data types and confirming its effectiveness through thorough experimentation. 4. SSCNN improves plain feature decomposition by incorporating a selection mechanism. This allows the model to identify fine-grained dependencies at each time step, which is essential for enhancing the accuracy of the decomposed structured components and, consequently, the overall prediction accuracy. 5. Extensive analysis has been performed to validate the method on existing benchmarks and compared with state-of-the-art methods. 6. Supplementary materials are satisfactory and provide explanation about the dataset and the implementation.

Weaknesses

1. Figures lack captions. 2. Include some limitations of the model as well. 3. second contribution and third one looks quite similar.

Questions

1. Please rewrite the contributions to make them clearer (2nd and 3rd). 2. Add descriptions in figure 2 and 3 3. In section D) Implications of Decomposition on Spatial-Temporal Correlations, please correct the captions of figures in temporal and spatial recovery.

Rating

7

Confidence

4

Soundness

4

Presentation

3

Contribution

3

Limitations

Authors have adequately addressed the limitations related to computational efficiency and capability of model.

Summary

The paper approaches the problem of long term time series forecasting (LTSF) using a compositional technique to reduce the model size without compromising the quality of solution. The proposed technique is a transformer based architecture with a lower number of parameters, and delivers similar performance as state of the art models for LTSF. The limitation of existing approaches, such as data patching, is that they fail to take into account the spatio-temporal dependencies, and end up with a blow up in the number of latent variables. This results only in a very small improvement even if the model size is increased substantially. The proposed technique in the paper is based on a inference step and an extrapolation step without any information loss. The paper evaluates the proposed approach, called SSCNN, with seven datasets, which has a combination of regular and volatile patterns. The baseline and state of the art approaches compared against include iTransformer, TimeMixer, and PatchTST. SSCNN consistently achieves the best scores, with respect to MSE and MAE. The paper also conducts ablation studies to show that each new component in the architecture is vital to the performance.

Strengths

The work studies an important and hard problem in time series forecasting which is the problem of efficient and accurate long term forecasting. Compositional techniques have been successful in other areas of AI including reinforcement learning, planning, and finite state controller synthesis. So, it makes sense to apply similar ideas in the space of long term time series forecasting.

Weaknesses

While the high level message is presented well, I found the details of the proposed method and experiments are hard to follow. A running example with the explanation of the new layers will be useful. The main contribution with respect to results is somewhat hard to grasp and align with the theoretical claims of the paper. Overall, I think there is room for improvement in the presentation of experimental results. I found some missing details in the experimental section that include: 1. Why is SSCNN missing Figure 3(a)? 2. What is the value of T_{out} in Figure 2? 3. What is the forward window size in Figure 3? In figure 2, it would be useful to move some of the methods to the appendix, and keep only the critical ones in the main body of the paper. Same is true for Figure 3. It is hard to go back and forth between figures 1, and figures2&3. Minor: 1. I would suggest providing some more details about the experimental results in point 3 of the contributions (lines 80-82) 2. Figure 3 is hard to read in print. 3. Having only one legend for all the subplots (Figure 2(a)-(d) and Figure 3(a)-(d)) will better than repeating the legends in all subplots.

Questions

1. Are there any non neural network based time series forecasting models which make use of compositionality? 2. Do any of the introduced layers (temporal, seasonal, short-term, etc) have similarities with any existing literature? What I mean is that, is the novelty in getting the layers to work together, or, also in defining the individual layers? 3. In order to better study the computational cost, can you share the total/average running time for each method for each dataset? 4. Why is SSCNN missing Figure 3(a)? 5. What is the value of T_{out} in Figure 2? 6. What is the forward window size in Figure 3?

Rating

6

Confidence

2

Soundness

2

Presentation

2

Contribution

3

Limitations

Yes, the paper has outlined the limitation of computational efficiency and provided some insight into how it can be improved in future work.

Thanks for responding to my queries

I would like to thank the authors for addressing my questions thoroughly. I have more clarity on the novelty aspect now.

Thank you, Authors!

Dear Authors, Thank you for your responses to Rviewers and your categorized, succinct overall response. On our end, we will collectively continue to review these. I hope the Reviewers' reviewers and discussion have helped strengthen your work and highlighted its significance and areas for improvement. Thank you for your submission and continued effort in the review process! Best, Area Chair xFv7