各式成本導向支持向量機之比較

Abstract

成本導向多重分類問題的應用不僅可見於許多方面，像是文件分類或是生物研究等，而且也是一個很熱門的研究領域。在這個問題中，為了使做出來的分類決策達到最低成本，有許多演算法被提出，然而至今卻沒有一個公平的平台去比較他們。而我們提供公正的環境，比較數種成本導向多重分類的演算法，配合支持向量機，並加以討論這些演算法的效能。除此之外，我們延伸了一個演算法到成本導向多重分類問題，並且推導出他的損失上限。我們的實驗指出，在成本導向多重分類問題中，這個延伸的演算法不僅比原來的演算法優越，而且也比一些成本導向多重分類問題的演算法好。

Multi-class cost-sensitive classification problem, which treats different cost for different types of misclassification, is currently attracting much research attention. Several promising meta-algorithms have been developed for this problem. Nevertheless, an empirical comparison of these algorithms has not been conducted. We would give a fair setup to compare them. In this thesis, we focus on the support vector machine (SVM) framework to solve the multi-class cost-sensitive classification problem. SVM is one of the most popular tools used in classification. In addition, we extend the all-together method in SVM formulation to the field of cost-sensitive problem. We also derive the theoretical guarantee of the extension. The experiment indicates this extension can get better performance than some existing regular classification algorithm and cost-sensitive classification algorithm in many cost-sensitive scenarios. The result also shows 'cost-sensitive one-versus-one','cost-sensitive one-sided regression' and this extension are more suitable than other cost-sensitive classification algorithms.