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

Summary

The paper introduces SOFTS (Series-cOre Fused Time Series forecaster), an efficient multivariate time series forecasting model that addresses the gap between channel independence and channel correlation in a novel way. By utilizing a centralized STAR (STar Aggregate-Redistribute) module, SOFTS creates a global core representation aggregated from all series, which is then redistributed and fused with individual series representations. This mechanism allows for efficient channel interaction while reducing dependence on the quality of each channel. The paper demonstrates SOFTS's superiority over state-of-the-art methods in terms of performance and computational complexity, and showcases the STAR module's adaptability across various forecasting models.

Strengths

- SOFTS proposes a unique approach to handling channel correlations in multivariate time series forecasting by employing a centralized STAR module, which aggregates and redistributes series representations efficiently. - The model's design and implementation are well-thought-out, and the empirical results show that SOFTS outperforms existing state-of-the-art methods. - The paper is well-written and structured, providing clear explanations of the STAR mechanism and its integration into SOFTS. - The proposed method has the potential to enhance forecasting accuracy and efficiency across various domains, making it a valuable addition to the field.

Weaknesses

- Additional experiments evaluating the robustness of SOFTS under varying conditions of distribution drift could provide deeper insight into its reliability.

Questions

- Has the author currently tried any other methods for aggregation and redistribution? - SOFTS is a linear based model, so how to handle non-linear time series prediction?

Rating

7

Confidence

4

Soundness

4

Presentation

4

Contribution

4

Limitations

- The authors have adequately addressed the technical limitations of their work within the paper.

Summary

This paper presents an efficient MLP-based model, the Series-cOre Fused Time Series forecaster (SOFTS). SOFTS incorporates a novel STar Aggregate-Redistribute (STAR) module to aggregate all series to form a global core representation, which is then dispatched and fused with individual series representations to facilitate channel interactions. The broad applicability of the STAR module across different forecasting models is also demonstrated empirically.

Strengths

1. The paper is well-written. 2. The proposed STAR module is designed as a core to aggregate and exchange information from the channels efficiently, which is a universal module and can replace the attention mechanism. 3. The experiments in Figure 6 are interesting and the results are impressive.

Weaknesses

The paper lacks significant innovation. The stochastic pooling itself is not novel in deep neural networks. A somewhat novel facet of the proposed model is to use the pooling to extract global representations. However, the idea of aggregate-and-dispatch the interactions between variables has already been studied in TimeXer [1]. [1] Wang, Y., Wu, H., Dong, J., Liu, Y., Qiu, Y., Zhang, H., ... & Long, M. (2024). Timexer: Empowering transformers for time series forecasting with exogenous variables. arXiv preprint arXiv:2402.19072.

Questions

1. It seems that all series share a core representation and MLP layers, but different variables will require different aspects of information. How can the model address these differences? 2. Please increase the look-back window length of the input series to 512 and 720, and compare the performance with other baseline models. 3. Please provide more results on abnormal channels, i.e. ETT, ECL, and Traffic dataset.

Rating

5

Confidence

4

Soundness

3

Presentation

3

Contribution

3

Limitations

yes

Summary

This paper proposes to use a global representation to capture the channel correlations for multivariate time series forecasting. Specifically, it uses stochastic pooling to get the global representation by aggregating representations of individual series and then concats the global representation and individual representations to reflect channel correlations for each series. Experiment results confirm that the proposal is much efficiency and achieves better performance than existing methods.

Strengths

1. This paper proposes an efficient method to capture the channel correlations for multivariate time series forecasting. 2. Extensive experiments are conducted to confirm the effectiveness of the proposal. 3. The paper is well-written and easy to understand in general.

Weaknesses

1. Some experimental results are unconvincing. -It is unclear why different datasets and metrics are used for different ablation studies, e.g., the datasets used in Table 3 and 4 are different, MAE is used in Figure 4 but MSE is used in other figures. -There is no statistical significance tests between the results of the proposal and baselines. -The results of Lookback Window Length 720 should be given as done in other papers. -It is better to give the training time for each methods as well. 2. Some places are not clearly described. -It is unclear the MLP and Linear operations are channel independent or dependent. I can guess they are channel independent, but it is not clearly described in Figure 1 and Section 3. -The sentence "... rely too heavily on the correlation to achieve satisfactory results under distribution drift" in the Abstract is not clearly explained. 3. Minor mistakes or typos: Embedding module is missing in Figure1; oi ∈ Rd' should be oi ∈ RC×d' in Algorithm 1.

Questions

Please refer to the weaknesses.

Rating

5

Confidence

5

Soundness

3

Presentation

3

Contribution

2

Limitations

Yes

Summary

The authors present a framework for modeling correlations between channels in a multivariate time series forecasting task. This framework concatenates each channel embedding with a ‘global core embedding’ which contains information from all channels in the lookback window. The authors present experiments that demonstrate the utility of this concept, both from a performance and efficiency perspective.

Strengths

Time series forecasting has been an important problem and it continues to grow with the advent of time series foundation models. To the best of my knowledge, it remains an open question for how to best enable multivariate time series forecasting, and this paper provides a conceptually reasonable approach. I believe the authors’ work is of broad interest.

Weaknesses

In my opinion, the paper would be improved with analysis and discussion of their results. There is little discussion beyond drawing attention to features of figures and tables, which misses an opportunity to explain why the authors believe they are observing such behavior. I am specifically interested in a discussion between iTransformer and SOFTS, which appear to be quite similar along many dimensions that SOFTS claims to be superior. Such a discussion will help guide potential readers through the considerations they should take into account when deciding which framework to implement on their own forecasting problems. I also recommend adding either a few sentences or a small figure that highlights the differences between PatchTST, transformer, and SOFTS — I see some details in the text of table 4, but I think making this information more prominent would be helpful and make the paper more self-contained.

Questions

1. Under what conditions should I choose SOFTS over iTransformer? Figure 3a makes it appear as though when there are a sufficiently large number of channels in the time series. Are there other considerations? Figures 3b and 6c show slight improvements but the difference doesn't seem large enough to discriminate between which framework would be better for a reader's specific time series forecasting efforts. 2. What are the limitations of SOFTS? Or what are the tradeoffs between using SOFTS vs other models? A few words of discussion in the conclusion would help improve the clarity of the paper.

Rating

7

Confidence

3

Soundness

3

Presentation

3

Contribution

3

Limitations

As is, there is no discussion of the limitations in the main text of the paper. To address this partially, section G of the appendix could be moved to the main text. However, the discussion of the limitations of SOFTS currently exists in a vacuum, lacking a comparison/contrast with other models mentioned in the paper, especially PatchTST and iTransformer. The authors should characterize and highlight the characteristics of datasets and inference tasks where SOFTS outperforms existing methods by wide margin.