<ul><li>Graph neural networks (GNNs) are limited by issues like constrained expressiveness, over-smoothing, and limited capacity to model long-range dependencies.</li><li>A new framework called Generative Graph Pattern Machine (G$^2$PM) has been introduced to overcome the limitations of message-passing in GNNs.</li><li>G$^2$PM uses generative Transformer pre-training over sequences of substructures to learn generalizable representations for graph instances.</li><li>On the ogbn-arxiv benchmark, G$^2$PM demonstrates strong scalability even with model sizes up to 60M parameters, outperforming prior generative approaches.</li></ul>

Scalable Graph Generative Modeling via Substructure Sequences

Discover more