DPLC: Dirichlet Process Guided Long-tail Clustering

The main challenge of long-tailed deep clustering is that imbalanced class frequencies cause head classes to dominate the partitioning of the representation space, which in turn submerges or distorts the cluster structures of tail classes. Existing methods typically alleviate this issue through heuristic head-tail partitioning or fixed auxiliary clustering. However, such approaches rely on manually specified discrete granularities and therefore struggle to capture the continuous latent structure of real-world long-tailed data. To address this limitation, we propose DPLC, a Dirichlet Process guided Long-tailed Clustering method that adaptively models the latent sub-cluster structure from a nonparametric Bayesian perspective. Specifically, DPLC periodically extracts soft latent sub-clusters from self-supervised embeddings and leverages them as data-driven rebalancing signals to mitigate the clustering bias induced by head-class dominance. By requiring neither a predefined number of auxiliary clusters nor explicit binary head-tail partitioning, DPLC enables more flexible modeling of complex long-tailed distributions. Extensive experiments show that DPLC consistently outperforms existing methods across multiple long-tailed clustering benchmarks and a wide range of imbalance ratios.