Boosted Personalized Page Rank Propagation for Graph Neural Prediction

Graph neural networks have recently achieved impressive success on many graph learning problems, including semi-supervised graph node classification. Previous message-passing algorithms are confronted with the dilemma of aggregating more neighbor’s information and avoiding over-smoothed node representations. The recently proposed personalized page rank propagation scheme separates the node prediction and information propagation into independent stages and achieves state-of-the-art performance on public datasets. However, the personalized page rank propagation scheme results in a computationally effective computation graph in model training. In this paper, we propose a boosted personalized page rank propagation scheme to reduce the model complexity and computational cost. The proposed algorithm is based on the principle of compressing multiple iterations in personalized page rank propagation into a one-step operation to simplifying the computation graph. Experiments on three public datasets reveal that the proposed algorithm reduces more than 50% training time, especially when the iteration number increases, without sacrificing the model’s prediction accuracy.