<p>Graph Neural Networks (GNNs) have achieved remarkable success in learning from structured data, yet most existing architectures struggle to model the intertwined semantics between nodes and edges-an issue especially prominent in heterogeneous graphs where relations are noisy, multi-typed, or semantically uneven. Such models usually learn node and edge features independently, leading to information disconnection and limited expressive power. To address this problem, we propose EdgeGFL, an <i>edge-empowered graph feature preference learning</i> framework that leverages multi-dimensional edge embeddings to guide node representation learning. By constructing relation-aware feature preference filters, EdgeGFL selectively amplifies informative feature channels and suppresses irrelevant ones during message passing, capturing non-local structural dependencies and fine-grained high-order semantics. This design bridges the gap between node- and edge-level learning while maintaining scalability. Extensive experiments on four real-world heterogeneous graphs (DBLP, ACM, IMDB, and Freebase) demonstrate that EdgeGFL consistently outperforms state-of-the-art GNN variants in node classification and clustering tasks. Beyond accuracy gains, the results highlight the importance of integrating edge semantics into GNN architectures for more robust and interpretable graph representation learning.</p>