Self-Supervised Symmetric Nonnegative Matrix Factorization

Symmetric nonnegative matrix factorization (SNMF) has demonstrated to be a powerful method for data clustering. However, SNMF is mathematically formulated as non-convex optimization problem, making it sensitive the initialization of variables. Inspired by ensemble clustering that aims seek better result from set results, we propose self-supervised (S 3 NMF), which capable boosting performance progressively taking advantage sensitivity characteristic SNMF, without relying on any additional information. Specifically, first perform repeatedly with random positive each time, leading multiple decomposed matrices. Then, rank quality resulting matrices adaptively learned weights, new similarity expected more discriminative reconstructed again. These two steps are iterated until stopping criterion/maximum number iterations achieved. We formulate S NMF constrained and provide an alternative algorithm solve theoretical convergence guaranteed. Extensive experimental results 10 commonly used benchmark datasets demonstrate significant our over 14 state-of-the-art methods in terms 5 quantitative metrics. The source code publicly available at https://github.com/jyh-learning/SSSNMF .