A probabilistic formulation NMF with group sparsity prior. Learning is variational Bayesian.
To be added.
For more details about the model, please refer to [1]. A short note: even though the formulation allows for inferring labels of unlabeled data, even simultaneousy with learning other model parameters, in practice it can be observed that the inferred labels often get stuck in local minima early on in the training stage. For this reason, in [1] this algorithm has been used to learn label-driven data representations, while for classification per se k-NN has been chosen.
20Newsgroups dataset. Left: group sparsity penalization; right: sparsity penalization.
If you have used this code, please cite the following paper, [1] I. Ivek: "Interpretable Low-Rank Document Representations with Label-Dependent Sparsity Patterns", Proceedings of DMNLP Workshop at ECML/PKDD / Cellier P., Charnois T., Hotho A., Matwin S., Moens M.F., Toussaint Y.; 2014
Published under GPL-3.0 License.