Data Augmentation for Inductive Link Prediction in Knowledge Graphs

Inductive link prediction in knowledge graphs aims to predict missing links between unseen entities, specifically relationships between entity pairs not observed during training. Current mainstream prediction methods mainly leverage the powerful representation ability of Graph Neural Networks (GNNs) to conduct inference on subgraphs around target nodes, aiming to uncover hidden semantic and structural patterns for link prediction. However, the information contained in subgraphs is relatively limited, making it hard to fully capture all patterns and structures of the original graph, especially with sparse data. These methods also inadequately consider the risk of overfitting in GNN training and the potential interference of noisy data, which limits the model's generalization and robustness. To address these issues, we propose the DAILP, a novel inductive knowledge graph embedding method. DAILP designs an enhanced subgraph strategy and specifically augments certain triples to enrich the knowledge graph's training-phase information. This enables the model to more effectively learn stable and rich graph structures and patterns, significantly enhancing its ability to adapt to complex new graphs. Moreover, to improve performance, we introduce adversarial training. By adding adversarial samples to the training data, we mitigate the impact of noisy data, enhance the model's robustness, prevent overfitting, and ensure stable performance with diverse inputs. Experimental results show that in inductive link prediction tasks, our model outperforms several baseline methods across different evaluation metrics.