STARK

Official PyTorch implementation of
STARK: Structure-Aware and Adaptive Representation Learning for Continual Knowledge Graph Embedding
(The Web Conference 2026, WWW 2026)

STARK addresses continual knowledge graph embedding, where new entities and relations are incrementally introduced over time. The framework adaptively allocates LoRA capacity based on structural importance and mitigates hubness via an adaptive TransE loss.

---

Overview

Continual knowledge graphs evolve as new facts, entities, and relations arrive over time. STARK is designed to efficiently adapt embeddings in this dynamic setting by jointly considering structural novelty and group-aware margin constraints.

Key features:

• Structural Novelty Prioritization (SNP) for identifying important new entities
• Adaptive TransE Loss (ATL) to alleviate hubness in many-to-one relations
• Efficient incremental updates via LoRA-based parameter adaptation

---

Requirements

Please refer to requirements.txt for the required dependencies.

---

Datasets

We evaluate STARK on continual knowledge graph benchmarks including:

• FB-CKGE
• ENTITY / RELATION / FACT / HYBRID

To Run the Code

To run STARK, simply execute the following command:

./main.sh

Acknowledgement

Our codebase is built upon the implementation of FastKGE. We thank the authors for their great work and for making the code publicly available.