An LLM-Guided Query-Aware Inference System for GNN Models on Large Knowledge Graphs

Expensive variable GNN inference: GNN inference on large knowledge graphs is computationally expensive and query complexity varies with target nodes and subgraph structure.

Non-query-adaptive acceleration: Existing acceleration methods create smaller models but do not adapt them to individual query structure or semantics.

Monolithic model loading: Existing systems store models as monolithic files that require full loading instead of retrieving only relevant model components and neighbors.

Excessive loading and redundant computation: Limitations in existing GNN inference acceleration lead to excessive data loading and redundant computation on large knowledge graphs.

Expensive variable GNN inference: GNN inference on large knowledge graphs is computationally expensive and query complexity varies with target nodes and subgraph structure.

KG-WISE task-driven inference: KG-WISE is a task-driven inference paradigm for large knowledge graphs.

Componentized GNN loading: KG-WISE decomposes trained GNN models into fine-grained components that can be partially loaded according to the queried subgraph structure.

LLM-generated query templates: KG-WISE uses LLMs to generate reusable query templates that extract semantically relevant subgraphs for each task.

Query-aware compact instantiation: KG-WISE enables query-aware and compact model instantiation for GNN inference.

KG-WISE task-driven inference: KG-WISE is a task-driven inference paradigm for large knowledge graphs.

Large-KG evaluation:

Faster inference:

Lower memory usage: