Maximum density minimum redundancy based hypergraph regularized support vector regression

Semi-supervised learning has attracted great attention in machine learning for it makes full use of labeled and unlabeled data for training. Most semi-supervised learning methods are not suitable for regression due to the data labels in regression are real-valued and smooth. In this paper, hypergraph instead of graph is utilized to represent the geometric structure of data. The manifold regularization term is constructed by calculating the hypergraph Laplacian and introduced into the regularization framework of kernel learning, a hypergraph regularized support vector regression (HGSVR) is proposed. Moreover, we propose a two-layer maximum density minimum redundancy method (MDMR) to pre-select initial labeled data, which fully considers the density and redundancy of data. The pre-select method is introduced into HGSVR and a second semi-supervised regression called MDMR-HGSVR is proposed. Experimental results on 9 UCI datasets show that HGSVR and MDMR-HGSVR outperform the other compared semi-supervised regression methods.

Automated summary

Semi-supervised learning is widely used in machine learning, but most methods are not suitable for regression. This paper proposes a hypergraph regularized support vector regression (HGSVR) method which utilizes the hypergraph to represent the geometric structure of data. Additionally, a two-layer maximum density minimum redundancy method (MDMR) is introduced to pre-select initial labeled data, and a second semi-supervised regression called MDMR-HGSVR is proposed. Experimental results on 9 UCI datasets demonstrate the superiority of HGSVR and MDMR-HGSVR over other compared methods.