Autoencoder-like non-negative matrix factorization with dual-graph constraints for multi-view clustering

Non-negative matrix factorization (NMF)-based multi-view data clustering has been widely used due to its simple formulation and strong interpretability.However, NMF-based multi-view clustering methods primarily focus on reconstructing the original data while neglecting the learning of low-dimensional representations, and emphasize learning the data manifold within the datasets while ignoring the learning of feature manifolds. To address these limitations, we propose a novel framework that combines autoencoder-like NMF with dual-graph constraints for multi-view clustering (ADGNMF). This approach unifies data representation learning and data reconstruction into a single framework, enhancing the learning of low-dimensional data representations. Additionally, to capture comprehensive information, we apply dual-graph constraints to both the data and feature manifolds. The algorithm employs an iterative updating strategy to optimize the objective function. Compared with several state-of-the-art multi-view clustering algorithms, ADGNMF has demonstrated superior performance across five key metrics on six public datasets.