Researchers have explored the expressivity of sparse graph neural networks (GNNs) within the framework of the Strong Expressive Lottery Ticket Hypothesis (SELTH). The study generalizes this hypothesis to multi-relational and temporal graph domains, introducing the concept of the relational Weisfeiler-Leman (RWL) algorithm. The findings indicate that sufficiently parameterized relational GNNs can contain sparse subnetworks that maintain 1-RWL expressivity, with a derived lower bound on the probability of achieving this through random pruning. The research also demonstrates that common temporal GNNs and cross-graph message passing schemes can be reformulated as RGNNs, inheriting these expressivity guarantees and connecting sparse RGNN expressivity to their optimization behavior. AI
IMPACT This research could lead to more efficient and expressive graph neural networks by better understanding the role of parameter sparsity.
RANK_REASON The cluster contains an academic paper detailing theoretical advancements in graph neural networks.
Read on Hugging Face Daily Papers →
- graph neural networks
- Hugging Face
- Panasonic RAW/RW2
- relational WL
- RGNNs
- Strong Expressive Lottery Ticket Hypothesis
- TGNNs
- Weisfeiler–Leman algorithm
- Nils Morten Kriege
- relational GNNs
- temporal GNNs
AI-generated summary · Google Gemini · from 2 sources. How we write summaries →
