Hypergraph regularized orthogonal non-negative tensor factorization for semi-supervised clustering

Numerous high dimensional data are collected from the internet of things (IoT) because of the rapid advancement of the sensors and data science. Due to cost issues, only a portion of the data can be labeled. Thus, it is significant to construct a semi-supervised dimensionality reduction method to learn from the IoT data. Non-negative tensor factorization (NTF) has been developed as an advanced approach for clustering, achieving superior performance by maintaining the spatial structure of high-dimensional data. However, most existing NTF methods are implemented in unsupervised tasks, resulting in neglect of supervisory information. In the paper, a hypergraph regularized orthogonal non-negative tensor factorization (HONTF) framework for semi-supervised clustering is proposed. Considering the Boolean structure of the ideal representation matrix which can be divided into labeled and unlabeled parts, the proposed method integrates the semi-supervised learning into the orthogonal NTF framework. Furthermore, hypergraph regularization is incorporated to capture multivariate relationships in low-dimensional space. Based on the HONTF framework, CP (CANDECOMP/PARAFAC) decomposition and Tucker decomposition are adopted to be the basic tensor factorization methods, respectively, exhibiting the scalability of the proposed framework. Extensive comparative experiments with six state-of-the-art semi-supervised methods on eight common real-world datasets and one IOT dataset demonstrate the superiority and effectiveness of the proposed methods.