Graph Neural Networks (GNNs) are successful at learning graph-structured data, but they struggle in heterogeneous graphs where connected nodes differ in features or class labels. To address these limitations, we propose GLANCE (Graph Logic Attention Network with Cluster Enhancement), a framework that enhances graph representation learning by integrating logic-based reasoning, dynamic graph refinement, and adaptive clustering. GLANCE combines a logic layer for interpretable and structured embeddings, multi-head attention-based edge pruning for noise removal, and a clustering mechanism for global pattern capture. Experiments on benchmark datasets from Cornell, Texas, and Wisconsin demonstrate that GLANCE achieves competitive performance, providing a robust and interpretable solution for heterogeneous graph scenarios.
