Learning attention-enhanced relation-sensitive knowledge graph embeddings with variable receptive fields for link prediction

Abstract

Knowledge graph embedding aims to predict missing links by learning representations of entities and relations. However, existing methods often struggle to effectively capture interaction features at multiple scales and to suppress redundant information among these multi-scale features, which hinders the model’s ability to focus on essential semantics. They also neglect the sensitivity of entities under specific relations, resulting in the lack of discriminability of entity representations. To tackle these problems, we propose a novel A ttention-enhanced R elation-sensitive V ariable receptive field-aware knowledge graph E mbedding method ( ARVE ). In this study, ARVE adopts a hybrid multi-scale grouped convolution module that extends the receptive fields of kernels in different forms to capture interaction features at multiple scales. Subsequently, ARVE introduces a feature points–convolutional kernels joint selection mechanism, which adaptively assigns weights to both convolutional kernels and specific feature locations, optimizing the adaptive fusion process of the aforementioned multi-scale features. Furthermore, a global attention block is employed to produce relation-sensitive entity representations by computing relevance scores between them, while modeling global dependencies. Experimental outcomes demonstrate that ARVE possesses a remarkable capability to capture information from local to global. It achieves notable performance across several benchmarks. In particular, on FB15k-237, ARVE improves MRR by 15.5% compared with the strongest graph neural network-based baseline DGATCL and by 6.6% compared with the overall strongest LLM-based baseline GLTW. Our code is available at https://github.com/JasmineMina07/ARVE .