Constructing a graduate employability knowledge graph with multi-source data fusion and graph neural network-based multi-task prediction

Addressing the limitations of existing employability prediction frameworks in cross-domain data integration and heterogeneous knowledge modelling, this study proposes a graduate employability knowledge graph construction framework grounded in multi-source heterogeneous data fusion and graph neural networks. A bidirectional confidence-scored alignment strategy is employed to integrate learning behaviour, market demand, and knowledge mastery data across three cross-domain sources, constructing a dynamic knowledge graph encompassing student, skill, and occupation entities across three hierarchical layers with seven semantically distinct relation types. An end-to-end multi-task prediction architecture incorporating basis decomposition parameter sharing and a dual-level attention mechanism is further developed upon this foundation. Experimental results validate the technical effectiveness of the proposed framework across skill mastery prediction, competency-level assessment, and job matching recommendation subtasks, as well as its representational stability under cross-dataset transfer conditions. The present study contributes a reproducible methodological framework for multi-source heterogeneous knowledge graph construction and graph-based multi-task prediction in educational contexts, with the findings offering a technical foundation for the design of intelligent career guidance systems and establishing a methodological basis for subsequent empirical investigations incorporating authentic graduate employment outcome data.