ResVGAE: Going Deeper with Residual Modules for Link Prediction

Learning-based feature extraction approaches have led to better performance in machine learning tasks, such as computer vision, machine translation, and object detection. Most real-world data sets have proven to be very successful in producing data representations that are successfully used in several tasks, such as fraud detection [1], recommendation systems [2], churn prediction [3] and predicting earthquakes using graph processes [4]. Graph neural networks (GNN) can efficiently exploit the relationship between data set instances in non-Euclidean space. Different variants of graph autoencoders, [5], [6], [7], [8], [9], have been very successful in capturing meaningful representations for node classification [10], link prediction [11] and graph classification [12] tasks.