Conditional Variational Graph Autoencoder for Air Quality Forecasting.

To control air pollution and mitigate its negative effect on health, it is of the utmost importance to have accurate real-time forecasting models. Existing deep-learning-based air quality forecasting models typically deploy temporal and-less often-spatial modules. Yet, data scarcity emerges as a real issue in this domain, a problem that can be solved by capturing the data distribution. In this work, we address data scarcity by proposing a novel conditional variational graph autoencoder. Our model is able to forecast air pollution by efficiently encoding the spatio-temporal correlations of the known data. Additionally, we leverage dynamic context data such as weather or satellite images to condition the model's behaviour. We formulate the problem as a context-aware graph-based matrix completion task and utilize street-level data from mobile stations. Experiments on real-world air quality datasets show the improved performance of our model with respect to state-of-the-art approaches.