ConvD: Attention Enhanced Dynamic Convolutional Embeddings for Knowledge Graph Completion (Extended Abstract)

作者：Wenbin Guo, Zhao Li, Xin Wang, Zirui Chen, Jun Zhao, Jianxin Li, Ye Yuan · 发表于：IEEE International Conference on Data Engineering · 年份：2026 · DOI：10.1109/icde65706.2026.00347

We propose ConvD, a dynamic convolutional embedding model that addresses knowledge graph incompleteness by enhancing relation-entity feature interactions. Unlike existing convolution-based KGC models that rely on fixed external kernels, ConvD reshapes relation embeddings into multiple internal dynamic kernels and integrates a knowledge-guided attention mechanism to adaptively weight their contributions. Experiments on multiple benchmarks demonstrate that ConvD achieves consistent performance gains of $\mathbf{3. 2 8} \boldsymbol{\%} \mathbf{- 1 4. 6 9 \%}$ over state-of-the-art methods while reducing parameters by $\mathbf{5 0. 6 6 \% - 8 5. 4 0 \%}$. This abstract is based on our TKDE paper which was accepted on June 10, 2025, and the e-copy is available at the DOI link: https://doi.org/10.1109/TKDE.2025.3582243.