基于Boosting框架的非稀疏多核学习方法

Non-sparse multiple kernel learning method based on Boosting framework

针对传统的分类器集成的每次迭代通常是将单个最优个体分类器集成到强分类器中,而其他可能有辅助作用的个体分类器被简单抛弃的问题,提出了一种基于Boosting框架的非稀疏多核学习方法 MKL-Boost。利用分类器集成学习的思想,每次迭代时,首先从训练集中选取一个训练子集,然后利用正则化非稀疏多核学习方法训练最优个体分类器,求得的个体分类器考虑了M个基本核的最优非稀疏线性凸组合,通过对核组合系数施加LP范数约束,一些好的核得以保留,从而保留了更多的有用特征信息,差的核将会被去掉,保证了有选择性的核融合,再将基于核组合的最优个体分类器集成到强分类器中。提出的算法既具有Boosting集成学习的优点,同时具有正则化非稀疏多核学习的优点,实验表明,相对于其他Boosting算法,MKL-Boost可以在较少的迭代次数内获得较高的分类精度。

Focus on the problem that the traditional classifier ensemble learning methods always integrated a single optimal classifier into the strong one, and the others, which maybe be useful to the optimal, were discarded simply in every Boosting iteration. This paper proposed a non-sparse multiple kernel learning method based on Boosting framework. At every iteration, firstly, this method selected a subset from the training dataset, then trained an optimal individual classifier by regularized non- sparse muhiple kernel learning method with this subset, which was obtained by optimizing the non-sparse combination of M basic kernels. It retained some good kernels and discarded the bad ones through imposing Lp-norm constrain on combination coefficients, and leaded to a selective kernel fusion and reserved more useful feature information. Lastly, these individual clas- sifiers were combined into the strong one. The proposed method has the advantages of ensemble learning methods as well as that of regularized non-sparse multiple kernel learning methods. Experiments show that it gains higher classification accuracy with smaller number of iterations compared with other Boosting methods.