Awesome

Taming Local Effects in Graph-based Spatiotemporal Forecasting (NeurIPS 2023 - pdf)

This folder contains the code for the reproducibility of the experiments presented in the paper "Taming Local Effects in Graph-based Spatiotemporal Forecasting" (NeurIPS 2023). The paper studies the interplay between globality and locality in graph-based spatiotemporal forecasting, proposing a framework to rationalize the practice of including trainable node embeddings in such architectures.

Authors: Andrea Cini, Ivan Marisca, Daniele Zambon, Cesare Alippi

Directory structure

The directory is structured as follows:

.
├── config/
├── lib/
├── conda_env.yaml
├── default_config.yaml
└── experiments/
    └── run_static_graph.py

Datasets

The datasets used in the experiments are provided by the tsl library. The CER-E dataset can be obtained for research purposes following the instructions at this link.

Configuration files

The config directory stores all the configuration files used to run the experiment. They are divided into subdirectories according to the experiment they refer to.

Requirements

To solve all dependencies, we recommend using Anaconda and the provided environment configuration by running the command:

conda env create -f conda_env.yml
conda activate taming-env

Experiments

The script used for the experiments in the paper is in the experiments folder.

run_static_graph.py is used to train and evaluate models on the datasets considered in the study. As an example, to run the TTS-IMP baseline with embeddings on the METR-LA dataset:
```
python experiments/run_static_graph.py config=benchmarks model=ttg_iso embedding=uniform dataset=la 
```