משה אליסוף

Graph Neural Networks (GNNs) achieved remarkable success on diverse tasks on graph-structured data, primarily by end-to-end learning. In this paper, we demonstrate that random weights can be surprisingly effective, achieving competitive performance with end-to-end training counterparts across various tasks and datasets. Specifically, we show that by replacing learnable weights with random weights, GNNs can retain strong predictive power, while significantly reducing training time by up to 6× and memory usage by up to 3× in the measured settings. Moreover, the random weights combined with our construction yield random graph propagation operators, which we show to reduce the problem of feature rank collapse in GNNs. These understandings and empirical results highlight random weights as a lightweight and efficient alternative, offering a compelling perspective on the design and training of GNN architectures when a suitable pretrained embedding is available.