Generative Hierarchical Mixture Density Network Clustering Model Based on von Mises-Fisher Distribution

This paper introduces a generative clustering model utilizing a hierarchical mixture density network framed within the context of the von Mises-Fisher (vMF) distribution. The proposed model constructs a dual-stage hierarchical mixture density network designed to map the multifaceted class features inherent in complex images, thereby establishing a potential distribution amongst the data. This evolution from a one-to-many to a many-to-many mapping paradigm empowers the model to extract deeper image category information through dual mappings, enhancing the efficacy of clustering analysis. Prior to training, we incorporated a preprocessing step to the data pipeline. This integration serves to preserve the coherence of the method and maintain the overall coupling integrity of the model, yielding a cohesive framework that supports joint training. The robustness of the proposed clustering model was evaluated using three benchmark image datasets and a medical image dataset. Comparative experimental results demonstrated that our model surpasses other clustering methods. Furthermore, the essential role of the mixture density network was substantiated through ablation studies.