SCINet: Time Series Modeling and Forecasting with Sample Convolution and Interaction

By Minhao Liu, Ailing Zeng, Mu-Hwa Chen, Zhijian Xu, Qiuxia Lai, Lingna Ma, Qiang Xu. Neural Information Processing Systems, 2021

851 citations recorded by the source corpus.

One unique property of time series is that the temporal relations are largely preserved after downsampling into two sub-sequences. By taking advantage of this property, we propose a novel neural network architecture that conducts sample convolution and interaction for temporal modeling and forecasting, named SCINet. Specifically, SCINet is a recursive downsample-convolve-interact architecture. In each layer, we use multiple convolutional filters to extract distinct yet valuable temporal features from the downsampled sub-sequences or features. By combining these rich features aggregated from multiple resolutions, SCINet effectively models time series with complex temporal dynamics. Experimental results show that SCINet achieves significant forecasting accuracy improvements over both existing convolutional models and Transformer-based solutions across various real-world time series forecasting datasets. Our codes and data are available at https://github.com/cure-lab/SCINet.

Subfields: Management Science and Operations Research, Signal Processing

View Paper: https://www.semanticscholar.org/paper/d9878cbb72b05f9b5ef0de7654209a25907c97b7

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Management Science and Operations Research

1. Qiang Xu
2. Ailing Zeng
3. Shiyu Wang
4. Donghao Luo
5. Xue Wang
6. Chenghao Liu
7. Doyen Sahoo
8. Gerald Woo

In Signal Processing

1. Donghao Luo
2. Xue Wang
3. Shiyu Wang
4. Lintao Ma
5. Zhijian Xu
6. Qiang Xu
7. Ailing Zeng
8. Haixu Wu

Citations in the corpus

• iTransformer: Inverted Transformers Are Effective for Time Series Forecasting (2023, methodology): “We use the same four public subsets (PEMS03, PEMS04, PEMS07, PEMS08) adopted in SCINet (Liu et al., 2022a).”
• Frequency-domain MLPs are More Effective Learners in Time Series Forecasting (2023, methodology): “…including Recurrent Neural Network-based methods (e.g., DeepAR [9], LSTNet [10]), Convolution Neural Network-based meth-ods (e.g., TCN [11], SCINet [12]), Transformer-based methods (e.g., Informer [13], Autoformer [14]), and Graph Neural Network-based methods (e.g., MTGNN [15], StemGNN [16], AGCRN…”
• FITS: Modeling Time Series with 10k Parameters (2023, methodology): “To effectively harness this data, recent research has given rise to sophisticated models and methodologies (Zhou et al., 2021; Liu et al., 2022a; Zeng et al., 2023; Nie et al., 2023; Zhang et al., 2022).”
• TimeMixer: Decomposable Multiscale Mixing for Time Series Forecasting (2024, methodology): “Baselines We compare TimeMixer with 15 baselines, which comprise the state-of-the-art long-term forecasting model PatchTST (2023) and advanced short-term forecasting models TimesNet (2023a) and SCINet (2022a), as well as other competitive models including Crossformer (2023), MICN Unified experiment ”
• ModernTCN: A Modern Pure Convolution Structure for General Time Series Analysis (2024, background): “Some previous convolution-based models (Wang et al., 2023; Liu et al., 2022a) try to bring convolution back to the arena of time series analysis.”
• Unified Training of Universal Time Series Forecasting Transformers (2024, methodology): “We report the Mean Squared Error (MSE) and MAE, comparing against six state-of-the-art baselines, iTransformer (Liu et al., 2023b), TimesNet (Wu et al., 2023), PatchTST, Crossformer (Zhang & Yan, 2023), TiDE, DLinear (Zeng et al., 2023), SCINet (Liu et al., 2022), and FEDformer (Zhou et al., 2022b).”