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

TL;DR

SCINet introduces a novel time series forecasting architecture that exploits preserved temporal relations under downsampling by recursively downsampling, convolving with multiple filters, and interacting between two sub-sequences. The core building block, SCI-Block, splits sequences into even/odd streams, applies distinct convolutions, and uses learned affine interactions to exchange information, organized into a binary-tree SCINet with residual fusion and a FC decoder. Stacked SCINet extends this with intermediate supervision to further enhance predictive accuracy, yielding state-of-the-art results across 11 real-world datasets for short-term, long-term, and spatial-temporal forecasting. Ablation studies confirm the necessity of interactive learning, residual connections, and multi-resolution design, while limitations highlight opportunities for probabilistic forecasting and integration of spatial relations in future work.

Abstract

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.

Figures (5)

• Figure 1: Existing sequence modeling architectures for time series forecasting.
• Figure 2: The overall architecture of Sample Convolution and Interaction Network (SCINet).
• Figure 3: The prediction results (Horizon = $48$) of SCINet, Autoformer, Informer, and TCN on randomly-selected sequences from ETTh1 dataset.
• Figure 4: Component analysis of SCINet on two datasets. Smaller values are better. See Section \ref{['sec:ab']}.
• Figure 5: The structure of $\phi$, $\rho$, $\psi$, and $\eta$.