The Hong Kong Baptist University Adapting Kernel-based Methods to Semi-supervised Learning: from Multi-class Svm to Spectral Analysis a Research Prospectus Submitted to the Thesis Committee for Pursuing the Degree of Master of Philosophy Department of Computer Science by Wu Zhili

This prospectus proposes a preliminary research topic about fusing the kernel-based SVM method and the similarity-based spectral clustering into a semi-supervised learning algorithm under the scope of learning from both labeled and unlabeled data. For the past nine months before the prospectus comes out, much effort has been put to extend the bloomed SVM to more practicable multi-class learning situations and to introduce the blooming spectral method to some semi-supervised learning scenarios, which, as a subordinate contributor, leads to two conference papers, one accepted, the other recently submitted. Furthermore, encouraged by some emerging progress in the field of semi-supervised learning from other pioneer researchers, my ongoing research is expected to be at a high feasibility rate, either in theory or practice. Hereafter, further study will be conducted to make the algorithm theoretically complete and practically realistic. Also, to validate the algorithm, besides testing on some standard datasets available in the machine learning community, comprehensive experimental results will be conducted to the task of text categorization.