Multitask Graph Neural Network for Knowledge Graph Link Prediction

Ye Wang, Jianhua Yang, Lijie Li, Jian Yao

Abstract: Predicting entities in knowledge graphs is a crucial research area, and convolutional neural networks (CNNs) have exhibited significant performance due to their ability to generate expressive feature embeddings. However, several existing methods in this field tend to disrupt entities and relational embeddings, disregarding the original translation characteristics in triples, leading to incomplete feature extraction.To address this issue and preserve the translation characteristics of triples, the present study introduces a novel representation technique, termed MultiGNN. The suggested approach uses a graph convolutional neural network for encoding and implements a parameter sharing technique. It employs a convolutional neural network and a translation model as decoders. The model’s parameter space is expanded to effectively integrate translation characteristics into the convolutional neural network, which allows it to capture these characteristics and enhance the model’s performance. The proposed method in this paper has demonstrated significant enhancements in several metrics on the public benchmark dataset when compared to the baseline method.