libsvm源代码注释+算法描述:svm_train

最新推荐文章于 2024-07-05 21:58:05 发布

qilin2016

最新推荐文章于 2024-07-05 21:58:05 发布

阅读量1.5w

点赞数

分类专栏： Machine Learning 文章标签： svm 源代码机器学习

本文链接：https://blog.csdn.net/zhoudi2010/article/details/50700332

版权

这篇博客主要聚焦于libsvm中解决C-SVM问题的solve_c_svc函数，该函数采用WSS3（一种使用WSS3的SMO分解方法）。首先选择最大化违反最优条件的特征对i和j作为工作集B。更新αi和αj，同时维护梯度。最终，svm_train训练后生成的模型文件包含了支持向量和其对应的权重。

摘要由CSDN通过智能技术生成

(I will try my best to make this note clearer. We mainly focus on solve_c_svc in this note)

We mainly focus on solve_c_svc in this note.

Our goal:
$\displaystyle \min_{\mathbf{d}_B}$ $\frac{1}{2}\mathbf{d}_{B}^{T}\nabla^2f(\alpha^k)\mathbf{d}_{B}+\nabla f(\alpha^k)_B^T\mathbf{d}_{B}$
s.t. $\mathbf{y}_B^T\mathbf{d}_B=0$

Core function of training: solve_c_svc.
It will call the function:

s.Solve(l, SVC_Q(*prob,*param,y), minus_ones, y, alpha, Cp, Cn, param->eps, si, param->shrinking);

Algotirhm 2 of WSS3 (An SMO-type decomposition method using WSS3)
After the initialization (trivial, we will discuss later), the first step is to select the working set $B=\{i,j\}$ . The chosen $i$ and $j$ most violate the optimally condition. Thus it is a natural choice . We called it “maximal violating pair”.
The pair satisfies in WSS1:
$i \in \arg \displaystyle \max_{t} \{ -y_t \nabla f(\alpha^k)_t | t \in I_{up}(\alpha^k) \}$
j∈argmint