<ul><li>The lottery ticket hypothesis (LTH) has been well-studied for convolutional neural networks but lacks theoretical validation for graph neural networks (GNNs).</li><li>This paper focuses on the expressivity of sparse subnetworks in GNNs, emphasizing their ability to differentiate non-isomorphic graphs as essential for identifying successful lottery tickets that maintain predictive performance.</li><li>The study establishes conditions where the expressivity of a sparsely initialized GNN matches that of the full network, especially concerning the Weisfeiler-Leman test. A Strong Expressive Lottery Ticket Hypothesis is proposed and proven in this context.</li><li>Increased expressivity in the initialization of GNNs is found to potentially speed up model convergence and enhance generalization. The research contributes new theoretical insights to both LTH and GNN studies, highlighting the significance of preserving expressivity in sparse GNNs, demonstrated through examples in drug discovery.</li></ul>

Weisfeiler and Leman Go Gambling: Why Expressive Lottery Tickets Win

Discover more