SVM是一种优化过的分类算法。和PLA的区别在于会选择“容错率”最大的hyperplane。
众所周知,模型的expressive power越强,generalization的可靠性越差。
SVM通过对“容错率”的考量中和了高纬度模型的风险。
SVM在数学上可以看做一种加入了regulizer 的回归法。
解SVM问题,可以只用标准QP solver。
Dual SVM problem has N variables to solve, as many as the number of the data points. The aim is to find out which vectors are support vectors.
In the procedure of solving a dual SVM problem, we need to calculate the inner product of every feature vector. It would be a lot of work with high dimention. Thus we can use kernel function to do these kinds of work to improve the efficiency.
We can choose the coefficient carefully to get more efficiency.
By kernel, we can use infinite dimention transformation of data.