Knowledge graphs about statistical data, such as the Eurostat Knowledge Graph (EurostatKG), present unique challenges for relation prediction due to their rich ontological schemas and strict semantic constraints. To address this, we reformulate relation prediction as a triple classification task and propose a unified, ontology-guided neural framework. Our approach integrates embedding-based models with graph-based semantic encoders (GCN, GAT), and introduces a hybrid loss function alongside a constraint-aware negative sampling strategy that respects ontology-defined domain and range axioms. Evaluated on EurostatKG, our method achieves strong results, including an F1-score of 0.96, PR-AUC of 0.99, Hits@5 of 0.618, and a Soft Hits@1 of 0.97. This work provides a principled integration of symbolic knowledge into neural models, while establishing a reproducible baseline for ontology-aware KG completion in complex, schema-rich domains.