SNAP-KG: Streaming Node Assignment via Projection for Knowledge Graph Entity Integration

Jui-Chien Lin; Mohammad Mohammadi Amiri; Oshani Seneviratne

Source-linked AI summary

SNAP-KG: Streaming Node Assignment via Projection for Knowledge Graph Entity Integration

Jui-Chien Lin, Mohammad Mohammadi Amiri, Oshani Seneviratne

arXiv:2608.25149v1Aug 25, 2026cs.LG

TL;DR

Streaming KG systems must assign newly arriving, graph-disconnected entities to communities before entity resolution and link prediction can search tractable candidate sets. SNAP-KG distills offline multi-view graph representations into a feature-only projector for immediate assignment without graph access or retraining. It matches retrain-level assignment quality at much lower inference cost and reduces candidate search across benchmark and production settings.

Problem
New entities lack graph connectivity, while transductive multi-view clustering cannot assign unseen nodes without retraining.
Method
SNAP-KG distills view-specific GNNs and a fused multi-view encoder into an MLP projector that assigns feature-only streaming entities to cluster centroids.
Results
The projector matches retrain-level cluster assignment quality with multiple orders-of-magnitude lower inference cost, while candidate search falls 62–75% on five benchmarks and 97% on OGB-WikiKG2.
Takeaways & Limitations
SNAP-KG provides inductive, graph-structural multi-view community assignment for scoping downstream entity resolution and link prediction.
Takeaways & Limitations
On MAG, scoping excludes 19.8% of ground-truth positive edges crossing cluster boundaries, and AUC-ROC falls from 96.43% to 92.21%.

Abstract

from arXiv · show

Knowledge graph (KG) construction pipelines must continuously integrate newly arriving entities into a growing graph. Unlike inserting triples between existing nodes, a newly arriving entity has no graph connectivity: it emerges from the acquisition phase as a raw feature vector and must be assigned to a semantic community before entity resolution and link prediction can operate over a tractable candidate set. Existing multi-view graph clustering methods exploit multiple relation types as structural views, but are transductive: they assume a fixed graph and cannot assign unseen entities without retraining. We propose SNAP-KG (Streaming Node Assignment via Projection for Knowledge Graph Entity Integration), a framework supporting graph-structural multi-view relational clustering and inductive inference for streaming entities. SNAP-KG trains a projector to map a new entity directly to the learned embedding space using only raw features, enabling immediate cluster assignment without graph access or model retraining. Experiments on five benchmark multi-view graph datasets and a production-scale KG of 2.4 million nodes demonstrate multiple orders-of-magnitude inference speedups over retraining-based approaches and competitive clustering quality. As a candidate scoping mechanism for downstream tasks, SNAP-KG achieves 62-75% candidate search reduction on the five benchmark datasets and 97% on OGB-WikiKG2 for entity resolution and link prediction.

1 Introduction

SNAP-KG addresses streaming KG integration by assigning disconnected entities to graph-structural multi-view communities without retraining or graph access. It distills offline graph-aware representations into an inductive projector that supports downstream candidate scoping.

Motivation: New entities arrive as raw feature vectors without edges or structural context, yet must be clustered before entity resolution and link prediction can restrict candidate sets.Without scoping, both tasks face linear-scale candidate search over existing nodes.
Research gap: Existing multi-view graph clustering is transductive, while existing inductive graph methods generally use single-view graphs or discard relational structure.SNAP-KG targets the missing combination of multi-view relational structure and inductive streaming assignment.
Approach: SNAP-KG trains view-specific GNNs and a multi-view Transformer encoder offline, then trains an MLP projector to reconstruct fused embeddings from raw features alone.At inference, the projector maps a new entity to the learned embedding space for nearest-centroid assignment.
Approach: The framework eliminates per-batch retraining and permits optional cluster refreshes without changing model weights.K-means is used in this work, but the framework is agnostic to the clustering algorithm.
Results: 62 %–75 % candidate search reduction is achieved across five benchmark datasets, alongside competitive clustering quality and multiple orders-of-magnitude lower inference cost.The evaluation also includes a 2.4M-node production KG.

2 Related Work

Prior work separates transductive multi-view graph clustering from inductive representation learning and streaming KG construction. SNAP-KG combines graph-structural multi-view learning with inductive assignment for unseen entities.

Multi-view graph clustering: Multi-view graph clustering methods achieve strong quality but require the complete graph and cannot extend embeddings to unseen nodes without full retraining.This transductive dependency conflicts with streaming ingestion.
Inductive learning: Inductive graph methods support unseen nodes without retraining, but prior GNN-to-MLP distillation evaluations use single-view architectures.Other inductive multi-view methods address feature modalities rather than graph-structural relation types.
Streaming KG construction: Streaming KG systems commonly lift structured streams into RDF triples or extract triples from unstructured text, whereas SNAP-KG operates downstream on entity representations.Its role is complementary to data-lifting and text-extraction systems.
Candidate generation: Traditional and neural blocking methods prune candidate spaces using node attributes but remain blind to relational structure.SNAP-KG instead uses graph-derived relational information during training while supporting feature-only inference.

3 Methodology

SNAP-KG models relation types as graph views, uses shared node features, and preprocesses neighborhoods adaptively before graph-aware training. The preprocessing changes training adjacency only, not stored KG semantics or inference.

Notation and node features: The method defines M relation-specific adjacency matrices over a shared node set and feature matrix X.Each view represents a distinct relation type, while node features provide the common input space.
Notation and node features: Node features use entity-specific surface forms, descriptions, literal values, and type labels without graph-derived statistics.The same feature extractor generates training features and streaming vectors.
Adaptive graph preprocessing: The preprocessing is motivated by noisy hub or co-occurrence edges whose aggregation can degrade representations, clustering, and projector learning.Feature similarity serves as a proxy for edge informativeness.
Adaptive graph preprocessing: Adaptive neighbor filtering ranks neighbors by feature-space similarity to remove weakly related connections before training.The neighborhood distribution determines whether a fixed neighbor budget or retention ratio is applied.
Adaptive graph preprocessing: Filtering is applied to filtered adjacency matrices throughout training only; it deletes no triples or predicates and is unused at inference.The stored KG and its semantics remain unchanged.

3.3 View-Specific Embedding Generation

SNAP-KG first computes view-specific node embeddings with separate GNNs, then combines their relational information through a multi-view encoder. The projector approximates this fused embedding process for streaming nodes.

View-specific embedding generation: Each view-specific GNN aggregates information from its filtered neighborhood to produce relation-specific node embeddings.Edges encode view-specific correlations and provide complementary local contextual information.
View-specific embedding generation: The projector network is trained to simulate the combined functionality of the view-specific GNNs and multi-view encoder.This enables efficient embedding generation for streaming nodes without reconstructing their graph neighborhoods.

3.4 Multi-View Aggregation via Encoder

A Transformer encoder fuses view-specific embeddings into one unified representation. Learnable positional encodings distinguish relation views, while an aggregation token attends across them.

A Transformer encoder aggregates the view-specific embeddings into a unified representation.
Learnable positional encoding p(m) distinguishes embeddings from different relation views.
A prepended aggregation token zagg attends to all views through self-attention.

3.5 Contrastive Learning

Contrastive learning encourages cluster-separable embeddings by defining positives from filtered neighborhoods across views and optimizing temperature-scaled similarities.

Contrastive learning encourages cluster-separable embeddings.
The positive neighborhood of node i is the union of its filtered neighborhoods across all views.
The node-level contrastive loss uses a temperature-scaled similarity metric.Similarity is defined as cosine similarity, with τ as the temperature parameter.

3.6 Projector Network

SNAP-KG adds an MLP projector that distills fused graph-aware embeddings from raw node features. This enables streaming-node embeddings without graph structure or retraining.

An MLP projector fϕ is trained to reconstruct fused embeddings from raw node features alone.It follows the GNN-to-MLP knowledge distillation paradigm and minimizes squared L2 reconstruction loss.
The projector enables inductive embedding for entities arriving without neighborhood information.
High-quality embeddings for streaming nodes can be generated without graph structure or retraining.

3.7 Joint Optimization

SNAP-KG jointly optimizes contrastive representation learning and projector reconstruction. The combined objective trains view-specific, cross-view, and inductive components together.

The full training objective combines the contrastive and projector losses.A balancing parameter λ controls the two objectives.
View-specific GNN encoders capture neighborhood structure within each relation type.
The Transformer integrates multi-view embeddings through cross-view attention.
The projector approximates the full GNN-Transformer pipeline from raw node features alone.

3.8 Streaming Inference and KG Construction Integration

SNAP-KG embeds streaming entities with a learned projector and assigns them to the nearest cluster centroid. These assignments restrict entity-resolution and link-prediction candidates to semantically coherent intra-cluster scopes while retaining task quality close to full-scan performance.

SNAP-KG embeds each new entity through a projector and assigns it to the nearest cluster centroid without retraining or graph reconstruction.The projector maps the raw feature vector directly into the learned embedding space.
O(N/K) comparisons replace O(N) entity-resolution scans when duplicate detection is restricted to intra-cluster candidates.K denotes the number of clusters.
Cluster assignments confine link prediction to intra-community pairs, providing a computationally tractable and semantically coherent scope for the expanding graph.The framework evaluates the resulting scope reduction for entity resolution and link prediction.
The induced cluster structure substantially reduces candidate search cost while keeping task quality close to the full-scan upper bound.

4 Experiment

SNAP-KG is evaluated for streaming inference efficiency, clustering quality, multi-view fusion, and downstream candidate scoping across benchmark graphs and a production-scale KG. The projector matches retrain-level assignment quality at far lower cost, while multi-view fusion improves clustering and scoping substantially reduces candidates, with boundary-related limitations for link prediction.

Five benchmark datasets—ACM, DBLP, IMDB, YELP, and MAG—are evaluated using node count, feature dimensionality, view count, and class count, with labels serving as clustering ground truth.
Streaming Inference: Speed and Quality: Milliseconds versus over 65 seconds per batch on MAG demonstrates multiple orders-of-magnitude faster projector inference than retraining-based approaches.SNAP-KG-Projector uses a feed-forward assignment and closely matches retrain-level ACC and F1 across the streaming evaluation.
Clustering Quality: Inductive vs. Transductive: SNAP-KG achieves competitive clustering across all datasets, including best MAG ACC of 68.20% and F1 of 68.35%.On ACM, DBLP, and YELP, ACC reaches 91.50%, 91.82%, and 91.20%, respectively, within 2% of the strongest transductive baselines.
Ablation Study: Multi-view fusion consistently outperforms single-view variants, raising ACM NMI from 41.28% to 71.08% and DBLP NMI from 5.02% to 74.03%.Consistent gains on YELP and MAG support the contribution of aggregating relation-specific views.
Entity Resolution & Link Prediction: SNAP-KG reduces candidate searches by 62–75% while preserving ER quality near full-scan performance, but cluster boundaries can exclude true LP edges and degrade link-prediction metrics.On MAG, 19.8% of true edges cross cluster boundaries; scoped AUC-ROC falls from 96.43% to 92.21%.

5 Conclusion

SNAP-KG supports streaming KG construction by combining graph-structural multi-view clustering with inductive assignment of unseen entities. It replaces retraining with lightweight projection while preserving competitive clustering and substantially reducing downstream candidate search.

SNAP-KG combines relational clustering and inductive inference for streaming KG construction.
A lightweight MLP projector maps raw node features into the learned embedding space through a single feed-forward pass, without graph access or retraining.
The projector matches retrain-level cluster assignment quality while achieving multiple orders-of-magnitude speedups over retraining-based baselines.
Experiments span five benchmark datasets and a 2.4M-node production KG, with 62–75% candidate search reduction on the benchmarks and 97% on OGB-WikiKG2.
Entity-resolution quality remains close to the full-scan upper bound, and DBLP-ACM evaluation tests generalization to unseen ACM entities under domain shift.