A Lightweight Sparse Interaction Network for Time Series Forecasting

Authors: Xu Zhang, Qitong Wang, Peng Wang, Wei Wang

AAAI 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive experiments on public datasets show that LSINet achieves both higher accuracy and better efficiency than advanced linear models and transformer models in TSF tasks.
Researcher Affiliation Academia 1Shanghai Key Laboratory of Data Science, School of Computer Science, Fudan University, Shanghai, China 2Universite Paris Cite, Paris, France EMAIL, EMAIL, EMAIL
Pseudocode No The paper describes the architecture and mechanisms in detail through text and diagrams (Figure 3), but it does not include any explicitly labeled pseudocode or algorithm blocks with structured steps.
Open Source Code No The paper does not provide any concrete statement about releasing source code for the methodology described, nor does it include any links to a code repository.
Open Datasets Yes We evaluate the performance of the proposed LSINet on 6 popular datasets, including Weather, Electricity, and 4 ETT datasets, covering a range of time steps (17420 to 69680) and variables (7 to 321) and have been widely employed in the literature for multivariate forecasting tasks (Nie et al. 2023; Wu et al. 2021; Zhou et al. 2022b).
Dataset Splits Yes All methods follow the same data loading parameters (e.g., train/val/test split ratio) as in (Nie et al. 2023).
Hardware Specification Yes Experiments are conducted on NVIDIA Ge Force RTX 3090 GPU on Py Torch.
Software Dependencies No The paper mentions 'Py Torch' as the framework used, but does not specify a version number for PyTorch or any other software dependencies.
Experiment Setup Yes For LSINet, the hidden size for patch embedding, position embedding (Eq. 2), and all used MLPs are fixed at 128. The multi-head h is fixed at 4. η {1, 3} is used for controlling the interval of using sparse regularization loss. The number of patch e N is fixed at 64 for sparse interaction learning. δ for controlling sparsity is fixed at 0.15, i.e., the sparse rate of C is 0.85. The number of stacked STI modules is fixed at 1 on all datasets. ... The learning rate is fixed at 1e-4. The batch size for 4 ETT datasets is fixed at 128 while for Weather and Electricity datasets are fixed at 64 and 32 respectively. All methods follow the same data loading parameters (e.g., train/val/test split ratio) as in (Nie et al. 2023). For each experiment, we independently ran 5 times with 5 different seeds for 30 epochs and reported the average metrics and standard deviations.