Graph-based Forecasting with Missing Data through Spatiotemporal Downsampling
Authors: Ivan Marisca, Cesare Alippi, Filippo Maria Bianchi
ICML 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | In this section, we report the results of the empirical analysis of our approach in different synthetic and real-world settings. We use mean absolute error (MAE) as the figure of merit, averaged over only valid observations (Eq. 2). The code to reproduce the experiments and the instructions to download and pre-process the datasets are available online. |
| Researcher Affiliation | Collaboration | 1IDSIA USI-SUPSI, Universit a della Svizzera italiana 2Politecnico di Milano 3Dept. of Mathematics and Statistics, Ui T the Arctic University of Norway 4NORCE, Norwegian Research Centre AS. |
| Pseudocode | No | No pseudocode or algorithm blocks are explicitly presented in the paper. |
| Open Source Code | Yes | The code to reproduce the experiments and the instructions to download and pre-process the datasets are available online.1https://github.com/marshka/hdtts |
| Open Datasets | Yes | Notably, all datasets used in our study are publicly available. ... Finally, the second new dataset introduced in this paper, named Eng RAD, contains 3 years of 5 historical weather variables sampled hourly at 487 grid points in England. The measurements are provided by open-meteo.com (Zippenfenig, 2023) and licensed under Attribution 4.0 International (CC BY 4.0). |
| Dataset Splits | Yes | For Graph MSO and PV-US, we divide the obtained windows sequentially into 70%/10%/20% splits for training, validation, and testing, respectively. For Eng RAD, containing 3 years of data, we use the year 2020 for testing and one week per month in the year 2019 for validation. We use all the samples that do not overlap with the validation and test sets to train the models. |
| Hardware Specification | Yes | We ran all experiments sequentially on a workstation running Ubuntu 20.04.6 LTS and equipped with one AMD Ryzen 9 5900X 12-core processor, 128GB 3200MHz DDR4 RAM, and two NVIDIA RTX A6000 GPU with 48 GB GDDR6 RAM. |
| Software Dependencies | No | The paper lists software libraries such as Py Torch, Py Torch Geometric, Torch Spatiotemporal, Py Torch Lightning, Hydra, Numpy, and Scikit-learn, along with citations to their respective papers. However, it does not specify the exact version numbers for these software components. |
| Experiment Setup | Yes | We used Adam W (Loshchilov & Hutter, 2019) as the optimizer with an initial learning rate of 0.001. We used the Reduce LROn Plateau scheduler of Pytorch that reduces the learning rate by a factor of 0.5 if no improvements are noticed after 10 epochs. We trained all models for 200 epochs of 300 randomly drawn mini-batches of 32 examples and stopped the training if the MAE computed on the validation set did not decrease after 30 epochs. ... We use an embedding size of dh = 64 for all hidden representations hi t and zi t. In the input encoder, we concatenate the input with node embeddings θi of size 32 and apply an affine transformation. For temporal processing, we use L = 4 layers with a downsampling factor d = 3. For spatial processing, we use K = 3 pooling layers. The decoder is an MLP with 2 hidden layers with 128 units. We use ELU as the activation function within the architecture. |