[Paper] Positional Encoder Graph Neural Networks for Geographic Data

JackKelly commented 2 years ago

This paper might be relevant: https://arxiv.org/abs/2111.10144

To quote the abstract:

Graph neural networks (GNNs) provide a powerful and scalable solution for modeling continuous spatial data. However, in the absence of further context on the geometric structure of the data, they often rely on Euclidean distances to construct the input graphs. This assumption can be improbable in many real-world settings, where the spatial structure is more complex and explicitly non-Euclidean (e.g., road networks). In this paper, we propose PE-GNN, a new framework that incorporates spatial context and correlation explicitly into the models. Building on recent advances in geospatial auxiliary task learning and semantic spatial embeddings, our proposed method (1) learns a context-aware vector encoding of the geographic coordinates and (2) predicts spatial autocorrelation in the data in parallel with the main task. On spatial regression tasks, we show the effectiveness of our approach, improving performance over different state-of-the-art GNN approaches. We also test our approach for spatial interpolation, i.e., spatial regression without node features, a task that GNNs are currently not competitive at. We observe that our approach not only vastly improves over the GNN baselines, but can match Gaussian processes, the most commonly utilized method for spatial interpolation problems. The code for this study can be accessed via: this https URL

JackKelly commented 2 years ago

Hmm, having now skim-read the paper, I'm not so sure it's entirely relevant to our work! But I could be wrong! I'll be honest, I'm not entirely sure I understood the paper :slightly_smiling_face: (although I've only skim read it!)

jacobbieker commented 2 years ago

From the skim read I did, I am not sure it is either! The Euclidian distance, or great circle distance does work for our data, as it is on a regular grid, and isn't particularly complex. Although it could maybe be more helpful in the assimilation network, which has slightly more complex inputs