Sparse Topic Classification on Large Sets of Documents via Non-negative Matrix Tri-factorization

This paper proposes a co-clustering framework for sparse topic classification, which incorporates orthogonality and sparse regularization via non-negative matrix tri-factorization The proposed model integrates two key components, namely norm regularization on both factor matrices to promote row sparsity, thereby enhancing noise robustness and facilitating the selection of informative features and representative topics; and linearly independent constraints on the data (row) and feature (column) cluster indicator matrices to promote distinct, non-overlapping clustering structures. An norm-based reconstruction loss is employed, and an effective algorithm is developed to obtain the solution. Numerical experiments on some benchmark datasets demonstrate that the proposed model and solution method achieve competitive performance in sparse topic classification, with improvements in clustering accuracy, sparsity, and interpretability compared with existing methods.