SNAP-KG Accelerates Knowledge Graph Entity Integration

Knowledge graph (KG) construction faces a continuous challenge: integrating new entities efficiently into an expanding graph. Unlike adding relationships between existing nodes, new entities arrive as raw feature vectors without prior graph connectivity, requiring assignment to a semantic community before further processing like entity resolution. Existing multi-view graph clustering methods are often transductive, meaning they require retraining when new entities appear, which is inefficient for streaming data. SNAP-KG (Streaming Node Assignment via Projection for Knowledge Graph Entity Integration) addresses this by providing an inductive framework. It trains a projector that maps a new entity's raw features directly into a learned embedding space. This allows for immediate cluster assignment without needing access to the full graph or retraining the model. Experiments on five benchmark datasets and a large-scale production KG demonstrate that SNAP-KG offers orders-of-magnitude faster inference compared to retraining-based methods while maintaining competitive clustering quality. It also significantly reduces candidate search for downstream tasks like entity resolution and link prediction.