MF-CLR: Multi-Frequency Contrastive Learning Representation for Time Series

Authors: Jufang Duan, Wei Zheng, Yangzhou Du, Wenfa Wu, Haipeng Jiang, Hongsheng Qi

ICML 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental To validate the effectiveness of MF-CLR, we conduct extensive experiments on five downstream tasks, including long-term and short-term forecasting, classification, anomaly detection and imputation. Experimental evidence shows that MF-CLR delivers a leading performance in all the downstream tasks and keeps consistent performance across different target dataset scales in the transfer learning scenario.
Researcher Affiliation Industry Jufang Duan 1 Wei Zheng 1 Yangzhou Du 1 Wenfa Wu 1 Haipeng Jiang 1 Hongsheng Qi 1 1Lenovo Research, Beijing, China.
Pseudocode Yes Algorithm 1 Hierarchical training.
Open Source Code Yes The source code is publicly available at https://github.com/duanjufang/MF-CLR.
Open Datasets Yes To balance between the general acceptance and suitability, we resample these public datasets by different frequencies along the feature dimension to meet the multi-frequency setting. Details about the public datasets preprocessing can be found in Appendix B.
Dataset Splits Yes The datasets are divided into training (70%), validation (10%) and test (20%) unless the given dataset has already been divided.
Hardware Specification No The paper does not explicitly describe the specific hardware (e.g., GPU models, CPU models, specific cloud instances) used for running the experiments.
Software Dependencies No The paper describes the model architecture and parameters but does not provide specific version numbers for any software dependencies (e.g., programming languages, libraries, frameworks).
Experiment Setup Yes For all the downstream tasks, we set the batch size to be 32, the initial learning rate to be 1E-3 with 70% decay every 10 steps, and α to be 1E-5. The encoder of MF-CLR contains 4 hidden dilated convolutional layers with feed-forward connections between consecutive blocks. The dilation is set to be 2i for the i-th layer with the hidden channel size of 64 and the kernel size of 3. Also, a 10% dropout is added to enhance robustness.