Graph Neural Networks: Application in Air Pollution Imputation

The issue of complete air pollution data has been receiving increasing attention in recent years. Imputing missing air pollution data is a crucial challenge in modern life. Existing methods typically rarely consider spatial correlation data attained by air pollution monitoring stations and use less additional information in each station such as meteorological data and geographic location data. In this paper, we propose a more advanced model for imputing missing data that leverages graph neural networks (GNNs). We formulate Matrix Completion on Graph structure by using Delaunay triangulations to construct clustering graph structure based on the geographic location data of different air pollution monitoring station, where the each edge between different nodes represents the similarity between two station. In terms of the neural network, we take the Graph Convolutional Networks (GCNs) to complete this task, which takes advantage of the connectivity information between nodes in a graph to update their feature representations. Furthermore, we compare it with conventional imputation models, such as convolutional neural network (CNN) and random forest. The lower Mean Squared Error (MSE) in our model, specifically 0.22 for SO2 and 0.0074 for CO, demonstrates the robustness of our method in imputing missing pollution data.