Angle-based Convolution Networks for Extracting Local Spatial Features

Convolutional Neural Networks (CNNs) have been a powerful feature extracting method for various machine learning problems, but it is limited to a regular grid shape spatial locations, e.g., pixels of images or GPS grids. However, due to physical or privacy preserving problems, spatio-temporal data in the real world often consist of irregular spatial locations. To overcome this limitation, we propose Angle-based Convolutional Networks (ACNs) that leverages the local feature extraction function of CNNs and the graph formulation of Graph Convolutional Neural Networks (GCNs). Our method considers angles among spatial locations and introduces coefficient weight for angle partitions that enables us to extract local features appears in spatio-temporal data without regular grids.