HYQNET: A Novel Neural-Symbolic Approach for Knowledge Graphs

The ability to answer complex first-order logic (FOL) queries on knowledge graphs is essential for automated reasoning. Symbolic methods offer interpretability but struggle with incomplete graphs. Neural approaches generalize better but lack transparency. HYQNET proposes itself as a neural-symbolic model that integrates the strengths of both.

Architecture and Functionality

HYQNET decomposes FOL queries into relation projections and logical operations over fuzzy sets, enhancing interpretability. To handle missing links in knowledge graphs, it uses a hyperbolic GNN-based approach for knowledge graph completion in hyperbolic space. This allows for the effective incorporation of the recursive query tree, preserving structural dependencies.

Advantages of Hyperbolic Space

The use of hyperbolic representations allows HYQNET to capture the hierarchical nature of logical projection reasoning more effectively than Euclidean-based approaches. Experiments on benchmark datasets demonstrate HYQNET's strong performance, highlighting the advantages of reasoning in hyperbolic space.

For those evaluating the deployment of reasoning models on knowledge graphs, there are trade-offs between on-premise and cloud approaches. AI-RADAR offers analytical frameworks on /llm-onpremise to evaluate these trade-offs.