Mining Potential Information for Multiclass Microarray Data Using Centroid-Based Dimension Reduction

In this paper, we propose a novel dimension reduction algorithm that implements an information fusion of Centroid-based feature selection and partial least squares (PLS) based feature extraction. This paper focuses on mining the potential information hidden in multiclass microarray data and interpreting the results provided by the potential information. Firstly, a centroid concept has been introduced to define the objective function of feature selection. In order to obtain the sparse solution, logistic regression with L1 regularization has been incorporated into the objective function. The Centroid-based feature selection is then proposed to solve the optimization problem. By using the One-Versus-All (OVA) techniques, the Centroid-based feature selection is extended to solve multiclass problems. Secondly, we perform feature important analysis on microarray data by Centroid-based feature selection to determine the information feature subset (biomarkers). Finally, PLS-based feature extraction is conducted on the selected feature subset to extract the features that best reflect the nature of classification. The proposed algorithm is compared with three state-of-the-art algorithms using eight multiclass microarray datasets. The experimental results demonstrate that the proposed algorithm performs effectively and is competitive. Furthermore, mining the potential information of the microarray dataset improves the interpretability of the results.