Multiview Deep Autoencoder-Inspired Layerwise Error-Correcting Non-Negative Matrix Factorization

Multiview Clustering (MVC) plays a crucial role in the holistic analysis of complex data by leveraging complementary information from multiple perspectives, a necessity in the era of big data. Non-negative Matrix Factorization (NMF)-based methods have demonstrated their effectiveness and broad applicability in clustering tasks, as they generate meaningful attribute distributions and cluster assignments. However, existing shallow NMF approaches fail to capture the hierarchical structures inherent in real-world data, while deep NMF ones overlook the accumulation of reconstruction errors across layers by solely focusing on a global loss function. To address these limitations, this study aims to develop a novel method that integrates an autoencoder-inspired structure into the deep NMF framework, incorporating layerwise error-correcting constraints. This approach can facilitate the extraction of hierarchical features while effectively mitigating reconstruction error accumulation in deep architectures. Additionally, repulsion-attraction manifold learning is incorporated at each layer to preserve intrinsic geometric structures within the data. The proposed model is evaluated on five real-world multiview datasets, with experimental results demonstrating its effectiveness in capturing hierarchical representations and improving clustering performance.