SFTeAST: Integrating Structure, Frequency and Temporal Signals for Temporal Knowledge Graph Completion

<title>Abstract</title> <p>Temporal Knowledge Graph Completion (TKGC) aims to infer missing temporal facts by jointly modeling structural dependencies among entities and the temporal evolution of relations, which is crucial for event prediction and dynamic reasoning. Existing methods typically rank candidate entities over the entire entity space during inference, leading to high computational cost and vulnerability to noisy samples. Moreover, in sparse and complex temporal scenarios, models are often sensitive to anomalous events, resulting in limited generalization and instability. Therefore, we propose a novel TKGC model named Structure Frequency and Temporal Aware Spiral Tensor Model (SFTeAST). Specifically, to efficiently capture structural information, we design a structural similarity module that estimates similarity scores between entities and relations by pre-training a one-layer multilayer perceptron (MLP). To capture the continuous temporal evolution of relations, we introduce a spiral complex temporal module inspired by mapping relations onto an Archimedean spiral timeline and encoding temporal quadruples into third-order tensors. On this basis, to enhance generalization and inference stability under sparse timestamp scenarios, we develop a score fusion mechanism for effective integration of structural and temporal information. Furthermore, to reduce the interference of noisy candidate entities during inference, we devise a historical frequency filtering module that constructs a historical frequency filtering matrix to efficiently eliminate low-probability entities based on event frequency statistics. Extensive experiments on public benchmark datasets demonstrate that SFTeAST achieves competitive performance and outperforms most existing TKGC methods across the majority of evaluation metrics. The experimental results validate the efficacy of integrating these components for TKGC.</p>