基于截止投票的SVM集成快速分类方法

Fast SVM Ensemble Classifying Approach Based on Cut-Off Voting

通过充分利用多个基分类器间的差异,集成分类器能够有效提高泛化精度,但是分类复杂度也随之增加。针对一类典型基于重采样和投票法的集成分类器,根据少数服从多数原则,在不影响分类精度的前提下给出了硬截止投票方法;针对基于B agg ing的SVM集成的特点,引入概率分析,分析根据集成中部分投票预测集成结果的损失概率,给出了基于损失概率的软截止投票方法,该方法可推广到其他基于重采样技术与投票法的集成分类系统。对一个人工数据集和两个UC I数据集的实验表明该方法在保证分类精度的前提下,大幅提高了分类速度。

Ensembles improve the classification accuracy by integrating diverse sub-classifiers with the increase of the classification complexity. Firstly, based on majority decision, this paper proposes a hard cut-off voting approach without affecting the classification accuracy. Then, according to the characteristics of Bagging based SVM ensemble, a probability analysis is introduced to estimate the loss probability of taking partial votes as the ensemble result, and a loss probability based soft cut-off voting approach is given. The approach can be applied to other resampling and voting based ensemble classification schemes. Experiments on an artificial dataset and two UCI datasets indicate that the proposed approach efficiently speeds up classifying with the high accuracy.