<![CDATA[Some researchers often collect features so the principal information does not lose. However, many features sometimes cause problems. The truth of analysis results will decrease because of the irrelevant or repetitive features. To overcome it, one of the solutions is feature selection. They are divided into two, namely supervised and unsupervised learning. In supervised, the feature selection can only be carried out on data containing labels. Meanwhile, in unsupervised, there are three approaches correlation, configuration, and variance. This study proposes an unsupervised feature selection by combining correlation and configuration using multidimensional scaling (MDS). The proposed algorithm is MDS-Clustering, which uses hierarchical and non-hierarchical clustering. The result of MDS-clustering is compared with the existing feature selection. There are three schemes in the comparison process, namely, 75\%, 50\%, and 25\% feature selected. The dataset used in this study is the UCI dataset. The validities used are the goodness-of-fit of the proximity matrix (GoFP) and the accuracy of the classification algorithm. The comparison results show that the feature selection proposed is certainly worth recommending as a new approach in the feature selection process. Besides, on certain data, the algorithm can outperform the existing feature selection.]]>
Copyrights © 2023
