Interpretable Sparse System Identification: Beyond Recent Deep Learning Techniques on Time-Series Prediction
Authors: Xiaoyi Liu, Duxin Chen, Wenjia Wei, Xia Zhu, Wenwu Yu
ICLR 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | The performance evaluation of the proposed method is conducted on comprehensive benchmark datasets, including ETT, Exchange, and ILI. Results reveal that our proposed method attains a significant overall improvement of more than 20% in accordance with the most recent state-of-the-art deep learning methodologies. We will show the results of global prediction and local rolling prediction. Furthermore, we will highlight the advantages of GLIP in terms of computational efficiency and long-term prediction compared to neural networks. The experimental settings can be found in Appendix. D. Table 1: Local rolling prediction errors in terms of MSE and MAE. Table 2: Results for ablation experiments on the Exchange dataset. |
| Researcher Affiliation | Collaboration | Xiaoyi Liu1, Duxin Chen1 , Wenjia Wei2, Xia Zhu2 & Wenwu Yu1 1 School of Mathematics, Southeast University, Nanjing 210096, China 2 Network Laboratory, 2012 Lab, Huawei Technologies Co. Ltd, Shenzhen 518129, China |
| Pseudocode | Yes | In Appendix. E, we provide pseudo-code for the entire algorithmic process. Algorithm 1: Global Local Identification and Prediction (GLIP) |
| Open Source Code | No | The paper provides links to the source code of baseline models (e.g., 'The source code is available at https://github.com/cure-lab/LTSF-Linear.' for DLinear), but it does not explicitly provide a link or statement about the open-sourcing of its own proposed GLIP model's code. |
| Open Datasets | Yes | We selected four benchmark datasets, namely ETTh, ETTm, Exchange, and ILI (Zeng et al. (2023)), for validation and testing purposes. These datasets are wildely used and consist of multivariate time series data that cover various real-world domains such as temperature, exchange, and disease. One can refer to (Wu et al. (2021); Zeng et al. (2023)) for detailed descriptions of these data. |
| Dataset Splits | Yes | Analogous to deep learning methodologies, the time series will be partitioned into distinct sets, which are training set, validation set, and testing set, with a partitioning ratio of 7:1:2. |
| Hardware Specification | No | The paper states: 'Each of the aforementioned experiments can be completed in a matter of seconds to minutes on a personal computer running on a CPU environment.' This statement does not provide specific hardware models (CPU model, GPU model if any, memory, etc.) to allow for replication. |
| Software Dependencies | No | The paper states it 'only uses a few machine learning techniques for implementation' and mentions specific versioned software for baselines (e.g., DLinear, FEDformer), but it does not list specific software dependencies with version numbers for its own proposed method (GLIP). |
| Experiment Setup | Yes | GLIP provides default tuning methods and manual tuning methods regarding hyperparameter tuning. default tuning methods include three hyperparameters tuning modes. The first mode involves keeping the hyperparameters unchanged. Most hyperparameters can be directly determined before model training and do not require tuning. For instance, the order of Θg is generally set to 2, and k1 = k2 = 0.8, γ = 1. The second mode is the conditional tuning method. It adjusts parameters based on the difference between the input data and global predictions, i.e., α, β. When the mean difference between global predictions and input exceeds half of the current input data s semi-range, indicating a significant global prediction error that may affect local predictions, we set α = 0.8, β = 0.9. Otherwise, we set α = 0.1, β = 0.9. The third mode is the adaptive tuning method. It is γ in l1 regularization. Considering that γ being too large or too small will not yield good predictive results, we define a feasible range for γ, such as [1e 7, 1e 1]. |