Core Concepts
Simplex-structured matrix factorization (SSMF) is a generalization of nonnegative matrix factorization that aims to find a low-rank matrix decomposition where the columns of one factor lie on the unit simplex. This paper proposes a novel approach that converts the standard minimum-volume SSMF problem in the primal space into a maximum-volume problem in the dual space, providing new insights and an efficient optimization scheme.
Abstract
The paper focuses on the concept of duality and uses the correspondence between primal and dual spaces to provide a new perspective on simplex-structured matrix factorization (SSMF). The main contributions are:
A new formulation for SSMF based on maximizing the volume of the dual simplex, which bridges the gap between two existing families of approaches: volume minimization and facet-based identification.
A study of the identifiability of the parameters with this new dual formulation, proving that it can recover the true factors under various conditions, including the sufficiently scattered condition (SSC) and separability.
An efficient optimization scheme based on block coordinate descent to solve the proposed dual volume maximization problem.
Numerical experiments on both synthetic and real-world datasets, showing that the proposed algorithm performs favorably compared to the state-of-the-art SSMF algorithms.
The paper first introduces the SSMF problem and reviews previous approaches, including separability, volume minimization, and facet-based identification. It then presents the new dual volume maximization formulation and analyzes its identifiability properties under different conditions. Finally, it describes the optimization algorithm and provides experimental results.