Overcomplete Knowledge Mining, Organization and Ensemble: A Multiple Kernel Support Vector Machine Approach

Although data mining techniques are made tremendous progress, "knowledge-poor" is still a large gap of the current data mining systems. Few researches notice the fact that useful knowledge not only is the final results of an intelligent classification, clustering or prediction algorithm, but also runs through the whole process of data mining in which much potential useful information is viewed as redundancy and discarded. In this paper, we propose a new framework: over complete knowledge mining, organization and ensemble to make fully used of redundant information, incorporate expert knowledge and enhance the robustness of the final decision. As a popular data mining tool, multiple kernel support vector machine (MK-SVM) is used to systematically carry out a series of data mining tasks in those three stages of the framework such as feature selection, classification, decision rule extraction, associated rule extraction, subclass discovery, multiple feature subset and decision rule set ensemble. This approach is applied for medical decision support and achieves good performance.