Abstract
In this paper, we proposed a new matrix-based feature selection method that used the hidden knowledge in the orthogonal features obtained from the two-dimensional principal component analysis feature extraction method with transfer learning to perform highly accurate unsupervised feature selection. We briefly named it the UFS2DPCA algorithm. In general, features can be classified as redundant, irrelevant, and relevant. Correlation is another concept of redundancy and perfectly correlated features are redundant. Accordingly, we first use the 2DPCA approach to directly extract the uncorrelated and orthogonal features from the 2-D image dataset. Uncorrelated and orthogonal features are among the best features. Next, we compute the correlation similarity between the main and extracted features. Finally, we make a weighted bipartite graph using two sets of features and the similarities between them, then we select the best features of the primary using the fast LAPJV algorithm. We evaluate the performance of the proposed UFS2DPCA algorithm on four well-known image datasets using K- Nearest Neighbor classifier. Results of comparative experiments between the proposed UFS2DPCA algorithm and eight state-of-the-art unsupervised feature selection algorithms show that the UFS2DPCA method outperforms other methods.