基于迹比优化的正交子空间支持向量机

Orthogonal subspace support vector machine via trace ratio optimization

针对多平面支持向量机机器学习算法的分类性能受特征数量限制的问题,提出一种正交子空间支持向量机(orthogonal subspace support vector machine,OSSVM).首先为每类数据寻找一个正交子空间,使得该类数据和其他类数据在子空间中的投影存在较大间隔;然后基于迹比优化提出求解OSSVM模型的迭代算法,再利用核方法将OSSVM扩展为非线性模型.实验结果验证了本文算法在数据分类中具有良好的泛化性能.

Multi-surface support vector machines have achieved great progress recently, however, the number of features limits their classification performance. In order to extract more features from data and improve the classification accuracy of multi-surface support vector machine, an orthogonal subspace support vector machine （OSSVM） is proposed. OSSVM seeks an orthogonal subspace for each class such that corresponding class has large margin from other classes after projection. An iterative algorithm is developed to solve OSSVM based on trace ratio optimization. OSSVM is also extended to do nonlinear classification with kernel method. Experimental results confirm that OSS- VM leads to good generalization performance in classification problems.