Rethinking Graph-Based Document Classification: Learning Data-Driven Structures Beyond Heuristic Approaches