Machine learning (ML) models behind the decisions that increasingly affect our lives may rely on historical data fraught with biases, which may lead to discrimination. These biases can manifest in two significant ways: data may reflect historical prejudices, leading to models that perpetuate these issues, or there may be insufficient data for certain groups, hindering the models’ ability to identify fair and accurate patterns. This paper introduces FaGGM, a fairness-aware generative graphical model designed to tackle these challenges. FaGGM incorporates a fairness regularization term into its graph structure learning algorithm, making the models based on these structures fairer. Additionally, it acts as a fair data generator and improves data representation for underrepresented groups. FaGGM is flexible and is compatible with most fairness definitions and score-based structure learning algorithms. Our experiments demonstrate that FaGGM mitigates bias and generates high-quality synthetic data, setting it apart from existing bias mitigation methods. ML models trained on this data show considerably increased fairness scores and smaller reductions in accuracy relative to comparable approaches.