<p>Label Distribution Learning (LDL) has attracted lots of attention in recent years. Studies show that different groups of samples may exhibit different label correlations, motivating the use of local label correlation in LDL. Based on the above ideas, recent studies introduced unified learning frameworks that jointly partition the training set into clusters and perform label manifold learning. However, real-world samples rarely conform to the rigid geometric assumptions, such as spherical or convex shapes, implicitly made by conventional clustering algorithms. This indicates that the underlying data manifold may exhibit complex non-Euclidean structures. Nevertheless, existing research studies the label distribution manifold in Euclidean space, ignoring its intrinsic geometry. To address this limitation, we propose a new LDL method called <i>Label Distribution Learning via Modeling Label Correlation on Gaussian Components</i>. First, the Gaussian Mixture Model (GMM) is employed to partition the training set. GMM creates a more robust partition by modeling the distribution of data and assigning data points based on posterior probabilities. Second, we model the local label manifold within each cluster on the Fisher–Rao geodesic distance, which is the appropriate metric for studying label correlation. Third, it jointly learns the GMM clusters and the local label manifold by restraining the influence of statistically unreliable labels, which may distort the manifold geometry. Comprehensive experiments on numerous benchmark datasets validate the superiority of our proposed model, highlighting the importance of respecting the intrinsic non-Euclidean geometry of the label space.</p>