Search for a command to run...
In the era of data-based decision-making, the Knowledge Graphs (KGs) are an indispensable tool to manage and exploit complex relational data. Below, we browse some of these reputation graphs and show the ranking of Robert Oppenheimer, using the expanding size and variety of these graphs as significant challenges for traditional reasoning methods. We propose a strong framework by employing GNNs for scalable and interpretable global reasoning over large KGs. We conduct an extensive study on seven state-of-the-art GNN designs—namely, GCN, GraphSAGE, GAT, GIN, SIGN, APPNP, and Cluster-GCN—that are representative, and cover a spectrum of trade-offs on scalability, expressiveness and compute. We conduct an extensive empirical study of these models on the OGBN-Arxiv benchmark, covering a wide range of factors such as classification accuracy, training time, inference time, memory efficiency, et al. SIGN and Cluster-GCN outperform the other models in terms of scalability and accuracy, which validate their applicability for practical deployment. SIGN's precomputation-based approach eliminates runtime message passing, while the partition-based graph training in Cluster-GCN allows mini-batch learning on extremely large graphs. APPNP consistently performs well with its specialized PageRank propagation technique, while GAT and GIN can perform strong local reasoning. Our modular evaluation setup is designed to be reproducible and extendable, which provides researchers with a reproducible benchmark to investigate GNN-based knowledge graph reasoning. This paper highlights several trade-offs on model complexity, accuracy and scalability, which are essential for choosing compatible GNN architectures, for different types of knowledge reasoning tasks. They open the door for further harmonization with logical rules, temporal dynamics and multi-modal data for addressing the higher-level problems of KG incompleteness, evolution and cross-domain reasoning.