DSparsE: Dynamic Sparse Embedding for Knowledge Graph Completion

Abstract

Addressing the incompleteness problem in knowledge graphs remains a significant challenge. Current graph completion methods, such as ComDensE (a representative of the fully connected network) and InteractE (a representative of the convolutional network), have certain limitations. Specifically, ComDensE is prone to overfitting and has constraints on network depth, while InteractE has limitations in feature interaction and interpretability. To overcome these drawbacks, we propose the Dynamic Sparse Embedding (DSparsE) model. This model employs sparse learning techniques, replacing the conventional dense layers with adaptable sparse ones. DSparsE incorporates a structure reminiscent of the Mixture of Experts (MoE) at the encoding stage and a residual structure at the decoding stage, which optimizes feature extraction and decoding without a significant increase of parameters. Comparative tests are evaluated on the FB15k-237 and WN18RR datasets. It is demonstrated that DSparsE outperforms both ComDensE and InteractE on FB15k-237 in terms of hits@1, with improvements of 2.3\% and 3.0\%, respectively.