Learning and reasoning over graphs is increasingly done by means of\nprobabilistic models, e.g. exponential random graph models, graph embedding\nmodels, and graph neural networks. When graphs are modeling relations between\npeople, however, they will inevitably reflect biases, prejudices, and other\nforms of inequity and inequality. An important challenge is thus to design\naccurate graph modeling approaches while guaranteeing fairness according to the\nspecific notion of fairness that the problem requires. Yet, past work on the\ntopic remains scarce, is limited to debiasing specific graph modeling methods,\nand often aims to ensure fairness in an indirect manner.\n We propose a generic approach applicable to most probabilistic graph modeling\napproaches. Specifically, we first define the class of fair graph models\ncorresponding to a chosen set of fairness criteria. Given this, we propose a\nfairness regularizer defined as the KL-divergence between the graph model and\nits I-projection onto the set of fair models. We demonstrate that using this\nfairness regularizer in combination with existing graph modeling approaches\nefficiently trades-off fairness with accuracy, whereas the state-of-the-art\nmodels can only make this trade-off for the fairness criterion that they were\nspecifically designed for.\n