基於不平衡資料進行支援向量機的高斯核函數之參數選擇

Abstract

現實中，我們時常看到類別不平衡的資料，即其中一個類別其樣本數相較於其他類別特別低，然而這類型的類別通常也是我們所感興趣的。傳統上的分類器由於未考慮類別不平衡的情況，因此面對這類型的資料，分類器容易把資料歸類於多數量類別 (majority class)。在本次研究中，針對不平衡資料我們採用支援向量機 (support vector machine, SVM ) 並使用高斯核函數 (Gaussian kernel) 進行二元分類，為了改進 SVM 的分類效果與整體效率，我們考慮兩個問題: 不平衡的類別與參數選擇。在第一個問題中，我們基於調整閾值的概念提出兩個新方法，分別為 ROC-SVM 與 b-SVM；而在第二個問題中，我們提出一個快速且簡單的方法來挑選 SVM 的參數，且該方法並未使用交叉驗證。本研究中使用真實資料與模擬資料來評估我們所提出的方法，而結果顯示在大部分的情況下，ROC-SVM與 b-SVM 表現皆優於先前的方法，且整體運算時間也有明顯下降。

Skewed class distributions often occur in a wide variety of real datasets in which at least one of classes has relatively small number of observations, usually the class of interest. A classifier induced by such an imbalanced dataset has typically high accuracy for majority class and poor prediction for the minority class. In this study, we focus on SVM classifier with Gaussian radial basis kernel for a binary classification problem. In order to take advantage of SVM and to achieve the best generalization ability with satisfying prediction power, we will address two important problems: imbalanced datasets and parameters selection. In the first problem, we proposed two novel adjustment methods, ROC-SVM and b-SVM, for adjusting the cutoff threshold of SVMs. In the second problem, we propose a fast and simple approach to optimize model parameters of SVMs without carrying out extensive k-folds cross validation. Extensive comparison with standard SVM and well-known existing methods are carried out to evaluate the performance of our proposed algorithms using simulated datasets and real datasets. The experimental results show that our proposed algorithms outperform over-sampling techniques and existing SVM-based solutions.