<ul><li>Research introduces Gradient Contribution (GC) method for efficient few-shot Graph Neural Architecture Search, addressing weight coupling problem in NAS.</li><li>GC method computes cosine similarity of gradient directions among modules to allocate modules to sub-supernets based on conflicting or similar directions.</li><li>Unified Graph Neural Architecture Search (UGAS) framework explores optimal combinations of Message Passing Neural Networks (MPNNs) and Graph Transformers (GTs).</li><li>Experimental results show that GC improves supernet partitioning quality and time efficiency, while architectures found by UGAS+GC outperform manually designed GNNs and existing NAS methods.</li></ul>

Towards Efficient Few-shot Graph Neural Architecture Search via Partitioning Gradient Contribution

Discover more