{and Virendra Singh N Sangeeth Nirmal Kumar Boran Chandramani Chaudhary}

Abstract

By leveraging graph structure, Graph Neural Networks (GNN) have emerged as a useful model for graph-baseddatasets. While it is widely assumed that GNNs outperform basic neural networks, recent research shows thatfor some datasets, neural networks outperform GNNs. Heterophily is one of the primary causes of GNNperformance degradation, and many models have been proposed to handle it. Furthermore, some intrinsicinformation in graph structure is often overlooked, such as edge direction. In this work, we propose GNNDLD,a model which exploits the edge direction and label distribution around a node in varying neighborhoods(hop-wise). We combine features from all layers to retain both low-pass frequency and high-pass frequencycomponents of a node because different layers of neural networks provide different types of information. Inaddition, to avoid oversmoothing, we decouple the node feature aggregation and transformation operations.By combining all of these concepts, we present a simple yet very efficient model. Experiments on six standardreal-world datasets show the superiority of GNNDLD over the state-of-the-art models in both homophily andheterophily.