CoMRes: Semi-Supervised Time Series Forecasting Utilizing Consensus Promotion of Multi-Resolution

Authors: Yunju Cho, Jay-Yoon Lee

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

Reproducibility Variable Result LLM Response
Research Type Experimental Experimental results show that our proposed method not only surpasses traditional supervised models in accuracy but also exhibits greater robustness when extending the prediction horizon. We evaluate our model on 8 popular datasets, including 4 ETT datasets (ETTh1, ETTh2, ETTm1, ETTm2), Weather, Traffic, Electricity, ILI. These datasets have been extensively utilized for benchmarking and publicly available on Wu et al. (2021) and the data statistics are shown in Appendix A.1.1. The experimental results, summarized in Table 1, indicate that our proposed model outperforms the baselines on most datasets.
Researcher Affiliation Academia Yunju Cho, Jay-Yoon Lee Graduate School of Data Science Seoul National University 1 Gwanak-ro, Gwanak-gu, Seoul 08826, Korea EMAIL
Pseudocode No The paper describes the methodology in prose and uses Figure 1 to illustrate the overall architecture of the model. There are no explicitly labeled pseudocode blocks or algorithms with structured, step-by-step procedures.
Open Source Code Yes Code is available at this repository: https://github.com/yjucho1/Co MRes
Open Datasets Yes We evaluate our model on 8 popular datasets, including 4 ETT datasets (ETTh1, ETTh2, ETTm1, ETTm2), Weather, Traffic, Electricity, ILI. These datasets have been extensively utilized for benchmarking and publicly available on Wu et al. (2021) and the data statistics are shown in Appendix A.1.1. Footnotes provide direct links to the datasets: ETT dataset (https://github.com/zhouhaoyi/ETDataset), Electricity dataset (https://archive.ics.uci.edu/ml/datasets/Electricity Load Diagrams20112014), ILI dataset (https://gis.cdc.gov/grasp/fluview/fluportaldashboard.html), Traffic dataset (https://pems.dot.ca.gov/), Weather dataset (https://www.bgc-jena.mpg.de/wetter/).
Dataset Splits Yes Each dataset was split into training, validation, and test sets. Mean squared error (MSE) and mean absolute error (MAE) were used as evaluation metrics, with each experiment repeated five times, and the average values reported (see Appendix A.4 for error bars). Following the setup from the Pathformer paper, all models used the same input and prediction lengths: L = 36 for the ILI dataset and L = 96 for the others. Prediction lengths were H {24, 36, 48, 60} for ILI and H {96, 192, 336, 720} for the others. More experimental details are in Appendix A.1. Table 5: Statistics of popular datasets for benchmark. Datasets ETTh1 ETTh2 ETTm1 ETTm2 Electricity ILI Traffic Weather Features 7 7 7 7 321 7 862 21 Timesteps 17,420 17,420 69,680 69,680 26,304 966 17,544 52,696 Split Ratio 6:2:2 6:2:2 6:2:2 6:2:2 7:1:2 7:1:2 7:1:2 7:1:2
Hardware Specification Yes All experiments are implemented in Py Torch and executed on a NVIDIA A6000 48GB GPU.
Software Dependencies No All experiments are implemented in Py Torch and executed on a NVIDIA A6000 48GB GPU. The model is trained using the Adam optimizer (Kingma, 2014) with a learning rate of 10 3. The default loss function is L1 Loss, and early stopping is applied after 10 epochs if no improvement is observed. While PyTorch is mentioned as the implementation framework, no specific version number for PyTorch or any other library is provided.
Experiment Setup Yes To ensure fair comparisons, we use the same hyperparameters as Pathformer in Chen et al. (2024). The model is trained using the Adam optimizer (Kingma, 2014) with a learning rate of 10 3. The default loss function is L1 Loss, and early stopping is applied after 10 epochs if no improvement is observed. ... all models used the same input and prediction lengths: L = 36 for the ILI dataset and L = 96 for the others. Prediction lengths were H {24, 36, 48, 60} for ILI and H {96, 192, 336, 720} for the others. ... The baseline model, MRes, includes four different patch sizes, same to those used in Pathformer repository7. These patch sizes are selected from the following options: {2, 3, 6, 12, 16, 24, 32}.