MixER: MLP-Mixer Knowledge Graph Embedding for Capturing Rich Entity-Relation Interactions in Link Prediction.

Knowledge graphs are often incomplete in practice, so link prediction becomes an important problem in developing many downstream applications. Therefore, many knowledge graph embedding models have been proposed to predict the missing links based on known facts. Convolutional neural networks (CNNs) play an essential role due to their excellent performance and parameter efficiency. Previous CNN-based models such as ConvE and KMAE use kernels to capture interactions between embeddings, yet they are limited in quantity. In this paper, we propose a novel neural network-based model named MixER to exploit more additional interactions effectively. Our model incorporates two types of multi-layer perceptions (i.e., channel-mixing and token-mixing), which extract spatial information and channel features. Hence, MixER can seize richer interactions and boost the link prediction performance. Furthermore, we investigate the characteristics of two core components that benefit in capturing additional interactions in diverse regions. Experimental results reveal that MixER outperforms state-of-the-art models in the branch of CNNs on three benchmark datasets.