A dynamic graph attention network with contrastive learning for knowledge graph completion

The objective of the knowledge graph completion (KGC) task is to improve the comprehensiveness and precision of knowledge graphs by forecasting the absent triples within them. In recent years, graph neural networks have become a core technique in KGC due to their advantages in processing graph-structured data. However, traditional graph neural network-based methods often perform poorly on sparse graphs and fail to detect the quality of neighbors during neighbor aggregation effectively. To tackle these challenges, we propose DGATCL, a dynamic graph attention network with contrastive learning for KGC. Specifically, we introduce a dynamic sampling strategy to adaptively select relevant neighbors during information propagation, effectively reducing the impact of noise and focusing on more relevant entities. In addition, we design a double-branch attention mechanism that jointly captures relation-aware edge importance and neighbor-wise contribution at the node level, enabling more effective aggregation of information. To further enhance the model’s performance, we incorporate a structure-aware contrastive learning strategy. By constructing positive-negative sample pairs and computing the contrastive loss, we effectively alleviate the graph’s sparsity problem. Experimental results demonstrate that our proposed method can learn high-quality neighborhood information in sparse graphs and performs better than state-of-the-art models in the KGC task. The code is available at https://github.com/xlz0517/DGATCL.