Hierarchical Clustering Support Vector Machines for Classifying Type-2 Diabetes Patients

摘要

Using a large national health database, we propose an enhanced SVM-based model called Hierarchical Clustering Support Vector Machine (HCSVM) that utilizes multiple levels of clusters to classify patients diagnosed with type-2 diabetes. Multiple HCSVMs are trained for clusters at different levels of the hierarchy. Some clusters at certain levels of the hierarchy capture more separable sample spaces than the others. As a result, HCSVMs at different levels may develop different classification capabilities. Since the locations of the superior SVMs are data dependent, the HCSVM model in this study takes advantage of an adaptive strategy to select the most suitable HCSVM for classifying the testing samples. This model solves the large data set problem inherent with the traditional single SVM model because the entire data set is partitioned into smaller and more homogenous clusters. Other approaches also use clustering and multiple SVM to solve the problem of large datasets. These approaches typical employed only one level of clusters. However, a single level of clusters may not provide an optimal partition of the sample space for SVM trainings. On the contrary, HCSVMs utilize multiple partitions available in a multilevel tree to capture a more separable sample space for SVM trainings. Compared with the traditional single SVM model and one-level multiple SVMs model, the HCSVM Model markedly improves the accuracy for classifying testing samples.