Text mining using non-negative matrix factorizations

This study involves a methodology for the automatic identification of semantic features and document clusters in a heterogeneous text collection. The methodology is based upon encoding the data using low rank nonnegative matrix factorization algorithms to preserve natural data non-negativity and thus avoid subtractive basis vector and encoding interactions present in techniques such as principal component analysis. Some existing non-negative matrix factorization techniques are reviewed and some new ones are proposed. Numerical experiments are reported on the use of a hybrid NMF algorithm to produce a parts-based approximation of a sparse term-by-document matrix. The resulting basis vectors and matrix projection can be used to identify underlying semantic features (topics) and document clusters of the corresponding text collection.