Node-Centric Knowledge Graph

ICLR 2026 Conference Submission20137 Authors

Keywords: Node-Centric Knowledge Graph, Local Neural Modules (MLP), Hierarchical Control Nodes, Reinforcement Learning, Explainable Reasoning, Knowledge Graphs, Information Retrieval, Knowledge Graph Construction

Abstract: Large-scale knowledge graphs often rely on static triplet-based structures that can be slow to update and query, struggle with both local and entity disambiguation, and require substantial global synchronization. To address these limitations, we introduce a novel node-centric neural knowledge graph architecture that balances localized inference with global consistency via a two-tier system. At the lower tier, each entity node has a lightweight multi-layer perceptron (MLP) for storing and reasoning over its immediate relationships, facilitating fast, autonomous updates and local disambiguation. At the higher tier, a set of control nodes employing reinforcement learning coordinates multi-hop queries and resolves conflicts that arise among local nodes. By delegating straightforward tasks to MLPs and escalating complex or contradictory cases to control nodes, the framework significantly reduces the overhead of global synchronization common in traditional knowledge-graph pipelines. This hierarchical design also enables efficient subgraph extraction: domain-specific slices can be lifted out, along with their associated control nodes, without losing essential reasoning capabilities. Empirically, our approach delivers large improvements over embedding and GNN baselines, with average gains of $\sim$40 MRR points and $\sim$47 Hits@10 points across FB15k-237 and WN18RR, establishing a new state-of-the-art in node-centric knowledge graph completion.

Primary Area: learning on graphs and other geometries & topologies

Submission Number: 20137