In wireless communications, optimizing the resource allocation requires the knowledge of the state of the channel. This is even more important in device-to-device communications, one typical use case in 5G/6G networks, where such knowledge is hard to obtain at reasonable signaling costs. In this paper, we study the use of graph-based machine learning methods to address this problem. That is to say, we learn to predict the channel state on a given link through measurements on other links, thus decreasing signaling overhead. In particular, we model the problem as a link-prediction one and we consider two representative approaches: Random Dot Product Graphs and Graph Neural Networks. The key point is that these methods consider the geometric structure underlying the data. They thus enable better generalization and require less training data than classic methods, as we show on our evaluation using a dataset of RSSI measurements of real-world Wi-Fi operating networks.