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

17 Jun 2021  ·  Minhao Liu, Ailing Zeng, Muxi Chen, Zhijian Xu, Qiuxia Lai, Lingna Ma, Qiang Xu ·

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.

PDF Abstract

Datasets


Results from the Paper


 Ranked #1 on Time Series Forecasting on ETTh1 (24) Multivariate (using extra training data)

     Get a GitHub badge
Task Dataset Model Metric Name Metric Value Global Rank Uses Extra
Training Data
Result Benchmark
Time Series Forecasting ETTh1 (168) Multivariate SCINet MSE 0.408 # 1
MAE 0.417 # 4
Time Series Forecasting ETTh1 (168) Univariate SCINet MSE 0.071 # 1
MAE 0.21 # 4
Time Series Forecasting ETTh1 (24) Multivariate SCINet MSE 0.3 # 1
MAE 0.342 # 4
Time Series Forecasting ETTh1 (24) Univariate SCINet MSE 0.029 # 1
MAE 0.127 # 5
Time Series Forecasting ETTh1 (336) Multivariate SCINet MSE 0.504 # 10
MAE 0.495 # 4
Time Series Forecasting ETTh1 (336) Univariate SCINet MSE 0.084 # 6
MAE 0.234 # 5
Time Series Forecasting ETTh1 (48) Multivariate SCINet MSE 0.361 # 1
MAE 0.388 # 4
Time Series Forecasting ETTh1 (48) Univariate SCINet MSE 0.041 # 1
MAE 0.154 # 4
Time Series Forecasting ETTh1 (720) Multivariate SCINet MSE 0.544 # 10
MAE 0.527 # 4
Time Series Forecasting ETTh1 (720) Univariate SCINet MSE 0.099 # 7
MAE 0.25 # 6
Time Series Forecasting ETTh2 (168) Multivariate SCINet MSE 0.342 # 1
MAE 0.38 # 4
Time Series Forecasting ETTh2 (168) Univariate SCINet MSE 0.158 # 2
MAE 0.311 # 3
Time Series Forecasting ETTh2 (24) Multivariate SCINet MSE 0.18 # 1
MAE 0.263 # 4
Time Series Forecasting ETTh2 (24) Univariate SCINet MSE 0.065 # 1
MAE 0.183 # 5
Time Series Forecasting ETTh2 (336) Multivariate SCINet MSE 0.365 # 6
MAE 0.409 # 6
Time Series Forecasting ETTh2 (336) Univariate SCINet MSE 0.166 # 1
MAE 0.329 # 9
Time Series Forecasting ETTh2 (48) Multivariate SCINet MSE 0.23 # 1
MAE 0.303 # 4
Time Series Forecasting ETTh2 (48) Univariate SCINet MSE 0.093 # 1
MAE 0.227 # 4
Time Series Forecasting ETTh2 (720) Multivariate SCINet MSE 0.475 # 8
MAE 0.488 # 5
Time Series Forecasting ETTh2 (720) Univariate SCINet MSE 0.286 # 11
MAE 0.429 # 10

Methods