Clustering Stability via Concept-based Nonnegative Matrix Factorization

One of the most important contributions of topic modeling is to accurately and the ectively discover and classify documents in a collection of texts by a number of clusters/topics. However, finding an appropriate number of topics is a particularly challenging model selection question. In this context, we introduce a new unsupervised conceptual stability framework to access the validity of a clustering solution. We integrate the proposed framework into nonnegative matrix factorization (NMF) to guide the selection of desired number of topics. Our model provides a exible way to enhance the interpretation of NMF for the effective clustering solutions. The work presented in this paper crosses the bridge between stability-based validation of clustering solutions and NMF in the context of unsupervised learning. We perform a thorough evaluation of our approach over a wide range of real-world datasets and compare it to current state-of-the-art which are two NMF-based approaches and four Latent Dirichlet Allocation (LDA) based models. the quantitative experimental results show that integrating such conceptual stability analysis into NMF can lead to significant improvements in the document clustering and information retrieval the ectiveness.