NED-GNN: Detecting and Dropping Noisy Edges in Graph Neural Networks

Graph neural networks have become the standard learning architectures in graph-based learning and achieve great progress in real-world tasks. Existing graph neural network methods are mostly based on message passing neural network(MPNN), which aggregates messages from neighbor nodes to update representations of target nodes. The framework follows the assumption of homophily that nodes linked by edges are similar and share the same labels. In the real world, the graphs can mostly follow the assumption. However, for nodes in the graph, the connections between nodes are not always connecting two similar nodes. We regard the edges as noisy edges. Such edges will introduce noise to message passing in the training process and hurt the performance of graph neural networks. To figure out the noisy edges and alleviate their influence, we propose the framework called Noisy Edge Dropping Graph Neural Network, short as NED-GNN. By evaluating the weights between sampled negative edges and existing edges for each node, NED-GNN detects and removes noisy edges. Extensive experiments are conducted on benchmark datasets and the promising performance compared with baseline methods indicates the effectiveness of our model.