Learning Schema Embeddings for Service Link Prediction: A Coupled Matrix-Tensor Factorization Approach

Schema information is increasingly crucial to improve service discovery, recommendation, and composition, addressing link sparsity and lack of explainability inherent in methods relying solely on triples. However, existing approaches predominantly utilize schema information as a rigid filtering mechanism, equivalent to fixed conditions that lack the capability to adaptively adjust based on model learning. This paper introduces a novel learnable schema-aware knowledge embedding framework that enhances service link prediction by synergizing entity, relation, and type embeddings through a coupled matrix-tensor factorization model. To our knowledge, this is the first approach that couples entity and relation embeddings to enable adaptive learning of <bold xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"><u>Schema</u></b> <bold xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink"><u>E</u></b>mbeddings (<bold xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">SchemaE</b>). Our framework is both expressive and easy to use, with the capability to generalize to existing bilinear models. Within this framework, we further propose the schema prompt method for embedding isolated nodes, which typically suffer from sparse relations or the absence of neighbors, leading to biased representation often overlooked in existing works. Despite embedding schema information, our model remains lightweight due to the introduction of a parameter-efficient strategy via type assists. We conduct extensive experiments on four public datasets, including comparisons with existing SOTA models, parameter analysis, performance validation on extended models, and visualization. The experimental results confirm the effectiveness and efficiency of the proposed model.