This study explored methods for predicting connections in a social graph under conditions of incomplete data. Classical algorithms were first applied, allowing a clearer understanding of the graph’s structure and the inherent limitations of traditional approaches. Their analysis highlighted the challenges posed by missing information and established a baseline for further experimentation. Subsequently, data were prepared for training a graph neural network (GNN), with key factors influencing its performance identified and optimized. The trained model showed strong adaptability and learning capacity, though its accuracy fell short of ideal prediction levels. A comparison with classical models revealed objective constraints that restrict the full predictive potential of GNNs when working with incomplete graphs. The results led to conclusions about the boundaries of applicability for both classical algorithms and GNNs in solving link prediction problems under limited information. Recommendations were proposed to enhance predictive accuracy, emphasizing the importance of resource allocation and model selection for specific tasks. Overall, the study demonstrates that while no model achieves perfection, both approaches can deliver practical, acceptable outcomes when applied thoughtfully. The findings highlight the ongoing relevance of social graph analysis and illustrate how machine learning techniques – particularly GNNs – offer valuable opportunities to improve the detection of hidden social connections in modern contexts.