Convolutional Neural Network-Based Entity-Specific Common Feature Aggregation for Knowledge Graph Embedding Learning

Deep learning models present impressive capability for automatic feature extraction, where common features-based aggregation have demonstrated valuable potential in improving the model performance on text classification, sentiment analysis, etc. However, leveraging entity-specific common feature aggregation for enhancing knowledge graph representation learning has not been fully explored yet, though diverse strategies in knowledge graph embedding models have been developed in recent years. This paper proposes an innovative Convolutional Neural Network-based Entity-specific Common Feature Aggregation strategy named CNN-ECFA. Besides, a new universal framework based on the CNN-ECFA strategy is introduced for knowledge graph embedding learning. Experiments are conducted on publicly-available standard datasets for a link prediction task including WN18RR, YAGO3-10 and NELL-995. Results show that the CNN-ECFA strategy outperforms the state-of-the-art feature projection strategies with average improvements of 0.6% and 0.7% of MRR and Hits@1 on all the datasets, demonstrating our CNN-ECFA strategy is more effective for knowledge graph embedding learning. In addition, our universal framework significantly outperforms a generalized relation learning framework on WN18RR and NELL-995 with average improvements of 1.7% and 1.9% on MRR and Hits@1. The source code is publicly available at https://github.com/peterhu95/ConvE-CNN-ECFA.