This chapter focuses on how to construct the communication network that underlies a federated learning (FL) system. It shows how the structure of the FL network affects both the statistical accuracy and computational efficiency of FL algorithms. The chapter begins by analyzing how network properties—such as node degrees and eigenvalues of the Laplacian matrix—influence convergence speed and per-iteration cost. It then discusses how to measure similarities and differences between local datasets, using tools like probabilistic modeling, gradient analysis, and neural embeddings. These measures form the basis for graph learning methods, where the FL network is inferred directly from data. The final section formulates graph construction as an optimization problem that selects edge weights to minimize dataset discrepancy while meeting connectivity and complexity constraints. Together, these sections provide a principled framework for designing or learning FL networks that balance accuracy, efficiency, and communication overhead.