Reproduce “SCINet: Time Series Modeling and Forecasting with Sample Convolution and Interaction” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Minhao Liu, Ailing Zeng, Muxi Chen, Zhijian Xu, Qiuxia Lai, Lingna Ma, Qiang Xu
SCINet is a multiresolution convolutional architecture for time-series forecasting that recursively separates sequences into even- and odd-indexed components, processes them with distinct filters, and restores lost context through cross-subsequence interaction. Across 11 dataset variants and forecasting settings, it generally outperforms recurrent, convolutional, Transformer, and graph-based baselines, while using reported O(T log T) computation instead of quadratic self-attention.
Key insights
- Even–odd downsampling is treated as a structural property of regularly sampled time series: trends and seasonal patterns can remain identifiable at lower temporal resolutions rather than being discarded by the split.
- Separate filters and cross-branch interaction let the model represent different dynamics in each subsequence and compensate for information lost through decomposition.
- Ablations indicate that interaction, unshared branch weights, residual connections, and the direct decoder all matter; deeper hierarchies help with longer inputs, whereas excessive stacking can promote overfitting.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.