题目：AGM-GNN: Adaptive Gated Memory based Graph Neural Network for air pollutants' propagation modeling
报告人：王硕
摘要：Due to the complex temporal and spatial dependencies, air quality prediction is often challenging. The existing work can be mainly categorized into 1D-time-series based method, in which the spatial correlation information between air monitoring stations was seldom used, and 2D-grid based method, which may introduce the inaccurate interpolation data. In this paper, we present AGM-GNN, an end-to-end graph neural network with adaptive gated memory, in air pollutants' propagation dynamics modeling and prediction. Based on the physical and dynamical characteristic of air pollutants, temporal correlation and concentration variation, which influenced by pollutants' accumulation, vertical diffusion and deposition, are explored using a customized Long Short-Term Memory (LSTM) model along with meteorology data. Also, the wind's information, stations' coordinates are encoded into the edges' feature to model the pollutants' horizontal transmission leveraging the neighborhood aggregation and message passing scheme of Graph Neural Network (GNN). Finally, we evaluate our model on the real-world observation data and it outperforms the other baselines.
报告人信息：王硕，彩云科技工程师