The Clustering-Based Initialization for Non-negative Matrix Factorization in the Feature Transformation of the High-Dimensional Text Categorization System: A Viewpoint of Term Vectors

摘要

Due to the non-negativity of the matrix factors, Non-negative Matrix Factorization (NMF) is favorable for transforming a high-dimensional original Terms-Documents matrix into a lower-dimensional semantic Concepts-Documents matrix in the text categorization. With the iterative nature of all NMF algorithms, the NMF matrix factors need initializing. In this paper, we propose a clustering-based method for initializing the NMF according to the term vectors instead of the document vectors as the previous researches.