In 5G networks, Cloud-Radio Access Network (C-RAN) architecture divides legacy base stations
into Radio Remote Heads (RRH) and Base Band Units (BBU). RRHs transmit and receive radio
signals, whereas BBUs process those signals. Thus, BBUs can be centralized in cloud processing
centers serving different groups of RRHs. An adequate allocation of RRHs to BBUs is essential
to guarantee C-RAN performance. With the latest advances in machine learning, this task can
be automatically addressed through supervised learning. This paper proposes a methodology for
allocating RRHs to BBUs in heterogeneous cellular networks relying on graph partitioning
through a graph neural network. Model performance is assessed over a dataset built with a radio
planning tool that implements a realistic Long-Term Evolution (LTE) heterogeneous network.
Results have shown that the proposed method improves performance of a patented state-of-theart
tool based on graph partitioning.