Graph Attention Network Based Dual-Channel Recommendation Model For Innovation And Entrepreneurship Education

Innovation and Entrepreneurship Education (IEE) in higher education institutions has created an urgent need for intelligent and personalized course recommendation systems capable of meeting the multidisciplinary and experiential essence of entrepreneurial learning. The existing Graph Neural Network (GNN) models to educational recommendation do not have mechanisms of dynamically integrating the semantics of structured knowledge graphs with real-time Internet of things (IoT) streams of behavior, which lead to poor performance on cold-start learners and a low variety of recommendations. Therefore, in this paper, a novel conceptual framework of Dual-Channel Graph Attention Network (DC-GAT) is introduced, which incorporates Knowledge Graph semantic reasoning with Narrowband IoT (NB-IoT) behavioral sensing for personalized IEE course recommendation. DC-GAT is based on the Collaborative Knowledge Graph (CKG) of IEE enriched with domain ontology objects such as courses, concepts, competencies, tools, and projects and enriched with NB-IoT-derived behavioral relationships representing patterns of learner presence, equipment interactions, and collaboration patterns. It uses a dual-channel architecture, i.e., (i) KG semantic channel based on TransR embeddings and multi-layer Graph Attention Network (mGAT) propagation of relation-aware attention coefficients, and (ii) IoT behavioral channel based on Temporal Convolutional Networks (TCN) with four residual blocks, which encode time-series behavioral features and pool attention, respectively. A new cross-attention fusion layer and adaptive softmax gating mechanism is used to dynamically avoid overattention and underattention to semantic and behavioral representations based on user-specific activity profile, automatically adding more weight on the behavioral channel in case of cold-start users.Extensive experiments on a real-world IEE dataset show that DC-GAT can achieve significant gains over the state-of-the-art baselines with a Precision (0.1034), Recall (0.1812), NDCG (0.1624), and F1 (0.1316), respectively, which are improvements of 20.8%, 19.0%, 17.1%, and 20.1% over the best baseline HetGNN-KGAT, respectively. The proposed framework serves as a new paradigm of the IoT-enabled educational recommendation and provides a practical implication in terms of the composition of the next-generation smart learning environment in IEE.