Browsing by Author "Safiye, Yaylaoğlu"

Now showing 1 - 1 of 1

Statistical learning in modeling interrelations among veriables: an aplication to metabolomics
(Bahçeşehir Üniversitesi Fen Bilimleri Enstitüsü, 2011-01-27) Safiye, Yaylaoğlu; Kurşun, Olcay
In some machine learning problems, large datasets are naturally organized into some groups of variables, which are called views in the literature. Views can be used to predict the same target variable, such as the class of a given sample, such as in Parallel Interacting Multi-view Learning (PIML). In this thesis, we deal with a more general case, where the views are designed to predict different but related target variables. The goal here is to develop a mechanism for incorparating the interrelations among the target variables into their predictions, along with the input variables in their own views. In this study, the predictions obtained from the training phase of each view are used as additional inputs to the next iteration. Iterations are repeated until the interactions between the views in consecutive iterations become stable. The interrelations and interactions among the views are modeled using Support Vector Machines (SVM) along with optimization-related methods such as leave-one-out cross-validation, k-fold crossvalidation, grid search and bootstrap resampling. The proposed method is compared with the classical regression implemented on single view in its application to a toy dataset and a real-world dataset of cancer (a metabolomics dataset obtained through nuclear magnetic resonance spectroscopy on tissue samples from healthy and cancerous human subjects in a study conducted by the biomedical engineering department at the University of North Carolina). The web of interrelations among the views might give insight to the clinicians in their research.