Knowledge graph embeddings based on 2d convolution and self-attention mechanisms for link prediction: Knowledge graph embeddings based on 2d convolution...

Link prediction refers to using existing facts in the knowledge graph to predict missing facts. This process can enhance the integrity of the knowledge graph and facilitate various downstream applications. However, existing link prediction models usually extract features only in a global or local scope, resulting in feature extraction being limited to a single scope. Additionally, to achieve optimal results, many models require increasing embedding dimensions and parameter numbers, which can lead to scalability issues when applied to large knowledge graphs. To address these issues, we propose a model that fuses the self-attention mechanism with 2D convolution for the link prediction task. The model utilizes a self-attention mechanism with numerous heads to capture feature interactions between entities and relations in the global scope. Furthermore, we innovatively introduce 2D convolution to capture feature interactions in the local scope. Results using FB15k-237 and WN18RR as standard link prediction benchmarks show that our fusion model has good comparable performance with current state-of-the-art models. In particular, compared to the ConvE model (which uses only 2D convolution), our proposed model achieves 13.7% and 14.7% improvement in MRR metrics, and compared to the SAttLE model (which uses only the self-attention mechanism) achieves 2.5% and 0.5% improvement in MRR metrics. Furthermore, due to the low-dimensional embedding of entities and relations, our proposed model has low complexity, good scalability, and thus can accomplish link prediction tasks on larger knowledge graphs in the real world.