Multiview feature augmented neural network for knowledge graph embedding

Dan Jiang

doi:10.1016/j.knosys.2022.109721

2022-11-01

PRE

DOI：10.1016/j.knosys.2022.109721

原文链接

原文求助

摘要

En 中文

Link prediction in knowledge graph embedding is a meaningful research topic. Knowledge graph embedding (KGE) focuses on the problem of predicting missing links based on triples. Neural networks have been the paradigm of choice in knowledge graph tasks. However, the general network KGE models lack attention to the spatial location connection between entities and relations and have a weakness in that they cannot capture global information. We found that multiview feature construction can obtain more feature information corresponding to entities and relations. Aggregation of multiple types of multiview spatial transform information is a critical issue. Therefore, we propose a knowledge graph embedding method called the multiview feature augmented neural network (MFAE), which involves three components: multiview spatial transform, feature fusion convolution and feature information augmentation. To precisely augment the fusion of the vector spatial transform, a feature augmented convolutional network with attentive information calculation is introduced as a triple prediction. The multiview spatial transform is combined with a feature augmented convolutional network, which captures global feature information, obtains multiple views of entity and relation information, and improves the effectiveness of KGE. We conduct extensive experiments on link prediction, the effect of different views and feature augmented neural network comparison on benchmark datasets such as FB15k-237 and WN18RR. Experiments show that MFAE delivers significant performance compared to the classical link prediction methods.(c) 2022 Elsevier B.V. All rights reserved.

Keyword:

Knowledge graph embedding

Multiview spatial transform

Feature fusion convolution

Feature information augmentation

Link prediction