Neural Nonnegative Matrix Factorization for Hierarchical Multilayer Topic Modeling

Abstract

We introduce a method for detecting latent hierarchical structure in data based on nonnegative matrix factorization. Datasets with hierarchical structure arise in a wide variety of fields, such as document classification, image processing, and bioinformatics. The proposed method, Neural NMF, recursively applies topic modeling in layers to discover overarching topics encompassing the lower-level features. We derive a backpropagation scheme that allows us to frame our method as a neural network. Numerical results on a synthetic dataset demonstrate that Neural NMF outperforms similar algorithms on a hierarchical classification task.