|
Abstract:
|
To construct a high dimensional biological network whose known for their complexity, several supervised learning methods have been proposed in the field of statistics. In fact, the problem arises when dealing with biological network structures in which nonconvex structures might reflect the true relations of the components rather than convex forms so that unsupervised learning procedures can outperform their supervised counterparts. In this study, it is aimed to show how an unsupervised learning technique can produce a correct biological network structure. In particular, it is done by constructing a dissimilarity matrix driven by a kernel-based strategy that implies using the automatically enlarged feature space rather than the original version of it. Furthermore, it is proposed to be applied entropy-based distance criterion in the selection of clusters unlike the generic methods that are known as radial basis techniques in order to capture nonconvex clusters. To conclude, the aim of this study is to find suitable dissimilarity matrix by an unsupervised way, then to detect the correct biological network structure with its correct clusters by a proposed entropy based criterion.
|
