AutoCoG: A Unified Data-Model Co-Search Framework for Graph Neural Networks

Neural architecture search (NAS) has demonstrated success in discovering promising architectures for vision or language modeling tasks, and it has recently been introduced to searching for graph neural networks (GNNs) as well. Despite the preliminary success, GNNs struggle in dealing with heterophily or low-homophily graphs where connected nodes may have different class labels and dissimilar features. To this end, we propose co-optimizing both the input graph topology and the model’s architecture topology simultaneously. That yields AutoCoG, the first unified data-model co-search NAS framework for GNNs. By defining a highly flexible data-model co-search space, AutoCoG is gracefully formulated as a principled bi-level optimization that can be end-to-end solved by the differentiable search methods. Experiments show AutoCoG achieves gains of up to 4% for Actor, 7.3% on average for Web datasets, 0.17% for CoAuthor-CS, and finally 5.4% for Wikipedia-Photo benchmarks. All codes will be released upon paper acceptance.