A Comparison of Strategies for Unbalance Sample Distribution in Support Vector Machine
Linkai Luo, Hong Peng, Q. Zhang, Cheng Yuan Lin · 2006
In applications of support vector machine (SVM), we often meet the problem that the distributions of two types of samples are unbalanced. Some authors had put forward some strategies to deal with this problem, but up to now comparisons of these strategies haven't been conducted. Comparing with four kinds of strategies which are reselecting sample, adjusting penalty weight, increasing dummy ordinary sample and increasing dummy support vector sample, this paper points out that the front two kinds of strategies which don't increase dummy sample are more suitable for practical problem than the back two kinds of strategies which increase dummy sample. In strategies that increase dummy sample, the strategy that increases the dummy support vector sample precedes to the strategy that increases dummy ordinary sample. Meanwhile, this paper also points out that the more unbalanced the sample distribution is, the better the effects of these strategies are. These conclusions have an instructive meaning and a reference value to choices of strategies for unbalance sample distribution problem