The analysis of large-scale bibliographic networks is essential for understanding scientific collaboration and knowledge dissemination. However, conventional machine learning models that rely on tabular data often fail to capture the rich relational structure inherent in scholarly networks, leading to suboptimal predictive performance. This paper presents a framework that integrates graph data science with machine learning to model these complex data more effectively. We construct a bibliographic knowledge graph and apply it to two fundamental tasks: author research domain classification (node classification) and citation recommendation (link prediction). Using engineering features derived from graph topology, including centrality measures, community structure, and topological indices, we significantly improve model performance. For author classification, a Random Forest model using these features achieved an ROC-AUC of 0.67, a 24% improvement over a baseline model. For citation recommendation, a LightGBM model achieved a ROC-AUC of 0.69, marking an improvement of 30%. These results demonstrate that leveraging the topological properties of bibliographic networks provides a substantial predictive lift, validating the efficacy of graph-based feature engineering to build more accurate and insightful models of scholarly activity.