In this paper, we consider a cell-free massive multiple-input multiple-output network that employs an access point (AP) switch on/off technique for energy saving. We propose a novel graph-based AP switch on/off strategy, that utilizes graph signal processing to infer a graph that captures the neighborhood of each AP. Additionally, we adapt two selection algorithms so that they utilize the learned graph to determine the set of active APs. Simulation results show that our graph-based scheme effectively learns the environmental characteristics and reduces the outage probability over the entire coverage area compared to the Euclidean distance-based scheme.