<p>Tag recommendations designed to automatically assign the most relevant tags to items to improve recommendation results. Research has shown metric learning that satisfies triangular inequalities is more likely to capture similarity relationships between users, items, and tags compared to inner product computation. However, existing metric learning-based tag recommendation methods face critical challenges: random negative sampling strategies result in excessive proximity between negative and positive samples in low-dimensional metric spaces, forming ambiguous decision boundaries. Such hard negative samples hinder models from distinguishing semantically similar candidate tags, leading to misclassification and performance degradation. To address these issues, this paper proposes MLGAN, a Metric Learning for Recommendation with Conditional Generative Adversarial Networks based Hard Negative Samples. The framework first embeds users, items, and tags into a low-dimensional embedding space through metric learning, establishing distance relationships among original samples. Subsequently, an adversarial metric learning architecture incorporates a hard sample generation module, where a conditional generative adversarial network (CGAN) synthesizes hard negative tags that users would unlikely select. This process reveals latent user-item-tag relationships from original samples, while a novel difficulty consistency loss enables precise control over the hardness level of generated negative tags through a two-pronged approach. Finally, adversarial training jointly optimizes metric space parameters for both original and generated samples, enhancing the model’s discriminative capability to improve recommendation accuracy and robustness. We conducted extensive experiments on both LastFm and Movielens datasets to analyze the impact of different parameters and intrinsically different components on the performance, and experimental results show that the MLGAN evaluation metrics outperform the most competitive baselines.</p>