Effective radio resource management is crucial for optimizing both current and future wireless communication networks. Significant research has focused on identifying optimal resource allocation policies, which is a challenging mathematical problem. Consequently, learning-based algorithms, in particular those based on Graph Neural Networks (GNNs), have emerged as a practical and effective solution. However, most studies have relied on synthetic data for testing, which necessarily offers a simplified version of the complex propagation phenomena present in real-life wireless systems. In this paper we address this gap by evaluating these algorithms on a real-life dataset. Our experiments demonstrate that these solutions are indeed viable for real-world applications. Furthermore, since data for training will necessarily stem from the past, we verify that non-stationarities of the real-life networks do not negatively impact the trained algorithm’s performance (i.e. time transferability). Finally, given that deployed networks are typically designed to follow certain pre-established patterns, we analyze and confirm that these algorithms not only perform well but they can also be transferred between networks while maintaining strong performance (i.e. spatial transferability). These results confirm the viability of GNNs as practical tools for resource allocation in real wireless networks.