A Hybrid Graph Contrastive Learning Approach with Negative Sampling for Knowledge Graph Completion

Knowledge Graph Completion (KGC) constitutes a pivotal task for predicting missing relationships in knowledge graphs through existing triple analysis. While Graph Neural Networks (GNNs) have demonstrated exceptional performance in KGC by exploiting graph structural patterns, their effectiveness is fundamentally constrained by inherent limitations such as over-smoothing – a phenomenon that causes entity representations to converge to indistinguishable states. This critical shortcoming impedes the model’s capacity to discern fine-grained semantic relationships within knowledge graphs. To overcome these challenges, we present SCCL-GNN, a novel negative sampling-enhanced graph contrastive learning framework that synergistically integrates contrastive learning (CL) with GNN architectures. Our methodology explicitly optimizes the representation space through two innovative negative sampling paradigms: (1) Cluster-aware negative sampling strategically selects intra-cluster negative samples to enhance discrimination of semantically proximate entities. (2) Scoring-driven hard negative mining employs entity scoring mechanisms to identify challenging negative pairs that maximize training signal efficacy. Through comprehensive experimentation on three benchmark datasets (FB15k-237, WN18RR, and Kinship), our framework significantly outperforms state-of-the-art baseline models. This work not only advances the frontier of KGC methodologies but also provides fundamental insights into multimodal learning paradigm design for graph representation learning.