Time-Series Forecasting for Out-of-Distribution Generalization Using Invariant Learning
Authors: Haoxin Liu, Harshavardhan Kamarthi, Lingkai Kong, Zhiyuan Zhao, Chao Zhang, B. Aditya Prakash
ICML 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We conduct extensive experiments on diverse datasets along with three advanced forecasting models ( backbones ). FOIL proves effectiveness by uniformly outperforming all baselines in better forecasting accuracy. and 5. Experiments and 5.2. Results and 5.3. Comparison with Distribution Shifts Methods and 5.4. Ablation Study. |
| Researcher Affiliation | Academia | 1School of Computational Science and Engineering, Georgia Institute of Technology, Atlanta, USA. Correspondence to: Haoxin Liu <hliu763@gatech.edu>, B. Aditya Prakash <badityap@cc.gatech.edu>. |
| Pseudocode | Yes | A. Algorithm; Algorithm 1 The training procedure of our FOIL. |
| Open Source Code | Yes | Reproducibility. All training data, testing data and code are available at: https://github.com/Aditya Lab/ FOIL. |
| Open Datasets | Yes | Datasets. We conduct experiments on four popular realworld datasets commonly used as benchmarks: the daily reported exchange rates dataset (Exchange) (Lai et al., 2018), the weekly reported ratios of patients seen with influenzalike illness dataset (ILI) (Kamarthi et al., 2021a), and two hourly reported electricity transformer temperature datasets (ETTh1 and ETTh2) (Zhou et al., 2021). and All training data, testing data and code are available at: https://github.com/Aditya Lab/ FOIL. |
| Dataset Splits | Yes | We mainly follow (Wu et al., 2022) to preprocess data, split datasets into train/validation/test sets and select the target variables. |
| Hardware Specification | No | The paper does not provide specific hardware details such as GPU/CPU models or memory specifications used for running experiments. |
| Software Dependencies | No | The paper mentions using 'Time Series Library' and 'Domain Bed' for implementations, but does not specify their version numbers or any other software dependencies with version numbers. |
| Experiment Setup | Yes | Regarding the horizon window length, we considered a range from short to long-term TSF tasks. For ETTh1, ETTh2, and Exchange, the lengths are [24, 48, 96, 168, 336, 720] with a fixed lookback window size of 96 and a consistent label window size of 48 for the decoder. For the weekly reported ILI, the lengths are [4, 8, 12, 16, 20, 24], representing the next one month to six months, with a fixed lookback window size of 36 and a consistent label window size of 18 for the decoder.. Note that, we lack the availability of suitable environment labels. We address this by dividing the training set into k, tuned from 2 to 10, equallength time segments to serve as predefined environment labels. |