Virtual-Coordinate Based Sampling and Embedding for Machine Learning with Graph Data

Graph embedding techniques remains the only alternative for Graph Neural Networks (GNNs) in machine learning for applications for which data in the form of graphs. GNNs, however, require substantial computational resources and are not able to provide certain advantages of conventional neural networks, such as transfer learning. The existing embedding based techniques, however, rely on random walk based strategies, which are not efficient in capturing the topology information, or use more structured strategies followed by feature engineering to achieve the required embedding. We propose VCNN (VirtualCoordinate based Neural Networks) technique that directly uses distances to an extremely small set (< 1 %) of random nodes (anchors) for network sampling and embedding, thus eliminating the feature engineering step. Experiments with two benchmark datasets show that a simple shallow neural network configuration using VC can achieve comparable performance to existing embedding-based techniques requiring deeper neural networks, thus also eliminating the need to experiment with multiple neural network configurations. This finding is particularly relevant for applications where computational efficiency is crucial. This research extends the machine learning toolbox for networks and data in the form of networks, highlighting the importance of strategic parameter use and model simplicity.