Structure-Guided Contrastive Signal Construction for Knowledge Graph Embedding

Learning highly discriminative entity representations for knowledge graphs remains a challenging task, particularly for contrastive learning methods that often fail to capture the rich structural context during sample construction. To address these problems, we propose SynConKG, a structure-aware contrastive learning framework that systematically integrates structural semantics throughout the entire learning process. SynConKG first constructs high-quality, multi-view entity representations, which guide a contrastive signal generation strategy based on topology-aware positive augmentation and semantic hard negative mining. To further improve training dynamics, a hardness-aware loss function adaptively emphasizes informative and challenging samples. Our experiments on the WN18RR and FB15k-237 benchmarks demonstrate the effectiveness of SynConKG, showing that it consistently achieves significant improvements over existing methods in link prediction tasks.