Time-FFM: Towards LM-Empowered Federated Foundation Model for Time Series Forecasting

Authors: Qingxiang Liu, Xu Liu, Chenghao Liu, Qingsong Wen, Yuxuan Liang

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

Reproducibility Variable Result LLM Response
Research Type Experimental Our comprehensive experiments indicate that TIME-FFM outperforms state-of-the-arts and promises effective few-shot and zero-shot forecaster.
Researcher Affiliation Collaboration Qingxiang Liu1,2 Xu Liu3 Chenghao Liu4 Qingsong Wen5 Yuxuan Liang1 1 The Hong Kong University of Science and Technology (Guangzhou) 2 Institute of Computing Technology Chinese Academy of Sciences 3 National University of Singapore 4 Salesforce AI Research 5 Squirrel AI
Pseudocode Yes The training procedure of TIME-FFM is elaborated in Algorithm 1.
Open Source Code Yes The code is available at https://github.com/City Mind Lab/Neur IPS24-Time-FFM/tree/main.
Open Datasets Yes We evaluate on 8 benchmark datasets from various domains: ETTh1, ETTh2, ETTm1, ETTm2, Electricity, Weather, Exchange, and ILI, which have been widely adopted for evaluating time series forecasting performance. Detailed introduction of implementation and datasets can be found in Appendix A.
Dataset Splits Yes Detailed descriptions of datasets. The dataset size is organized in (training, validation, test). (8545, 2881, 2881) [for ETTh1] and The round with lowest validation value serves as the optimal round, and then the corresponding model is used for test.
Hardware Specification Yes All models are implemented on PyTorch with all experiments conducted on NVIDIA A100-80G GPUs.
Software Dependencies No All models are implemented on PyTorch with all experiments conducted on NVIDIA A100-80G GPUs. The Adam optimizer with the initial learning rate of 10 4 is adopted in the training process.
Experiment Setup Yes The Adam optimizer with the initial learning rate of 10 4 is adopted in the training process. The lookback window Li is set to 36 for the ILI dataset, and 96 for the others. The future prediction window Fi is set to {24, 36, 48, 60} for the ILI dataset, and {96, 192, 336, 720} for other ones. We adopt the pretrained GPT2-backbone of the first 6 layers as the LM encoder. The local epoch E is set to 1 for all domains. The global round number T is set to 100. V , M, P and H are set to 100, 12, 16, and 8 respectively for all domains. Si is set to 4 for the ILI dataset, and 16 for other ones.