| Summary: | 博士 === 國立成功大學 === 電機工程學系碩博士班 === 97 === Support vector clustering (SVC) has been widely researched in both theoretical development and practical applications due to its outstanding features—arbitrary-shaped cluster representations. SVC involves three main steps: 1) finding the hyper-sphere by solving the Wolfe dual optimization problem, 2) identifying the clusters by labeling the data points with cluster labels, and 3) searching a satisfactory clustering outcome by tuning kernel parameters. These three steps make using SVC to process large datasets inefficient and time-consuming. Based on the above problems, an efficient data preprocessing procedure is first proposed to eliminate insignificant data points from training datasets without significantly affecting the final cluster configuration. Since the size of dataset is reduced, the computational burden for solving the optimization problem as well as cluster labeling can be greatly decreased. Next, an optimal parameter search method is proposed to find the suitable parameter of kernel functions and soft-margin constant of Lagrangian functions of SVC. This dissertation enables SVC to identify optimal cluster configurations with a less number of executions. The applicability of the proposed approach to real-world applications is validated through music emotion classification problems. The classification process of the music samples includes feature extraction, features selection, feature transformation, and classification. Each of these processes possesses its own significant and plays an important role in music emotion classification. Finally, the effectiveness of the proposed approach is successfully validated by computer simulations on benchmark datasets and real-world applications.
|
|---|
