Hierarchical Dynamic Self-Supervised Learning for Robust Deep Non-negative Matrix Factorization in clustering tasks

Deep Non-negative Matrix Factorization (DMF) holds immense potential for learning hierarchical data representations. However, the commonly used pre-training and fine-tuning paradigm may suffer from an optimization inconsistency: the reconstruction-driven fine-tuning stage can degrade the hierarchical representations learned during pre-training. To resolve this, we propose Hierarchical Dynamic Self-Supervised Learning (HDSSL), a framework for training DMF models with intermediate self-supervised guidance. Unlike traditional methods that rely solely on a single global objective, HDSSL introduces a dynamic self-supervision mechanism that acts as a hierarchical structural regularizer. It iteratively refines target representations by projecting global data into the current latent space and updates network parameters via a hierarchical gradient aggregation algorithm. This approach explicitly injects consistent intermediate guidance, effectively mitigating optimization conflicts and reducing the risk of representation degradation during optimization. Extensive experiments on benchmark datasets demonstrate that HDSSL achieves consistently competitive clustering performance and favorable optimization behavior across different settings. Moreover, it shows faster convergence and improved stability in our experiments, while eliminating the need for layer-wise pre-training.