训练SVM的参数是什么 [英] What are the parameters for training SVM

问题描述

我正在使用该库

它指出 C 定义了错误分类和边际之间的权衡.必须根据您的数据选择足够大的值.您还将在这里看到 eps> 0 参数.这可能是您的 tolerance 参数，并定义了目标函数中由 C 参数加权的的误差.

对于内核参数，请看一下 SVM 的双重问题:

您会看到术语 K(x_i，x_j).这称为内核功能.该功能允许SVM学习非线性决策边界.因此，如果您的数据不是线性可分割的，则可以使用这样的函数将数据(实际上是 dot-product )转换为更高维度的特征空间，以将其分离.只需看一下本指南，它将教您有关SVM培训过程的基础知识和一些最佳实践:

https://www.csie.ntu.edu.tw/~cjlin/papers/guide/guide.pdf

I'm doing software using machine learning with SVM using this library https://www.npmjs.com/package/machine_learning

according to the example of SVM:

svm.train({
    C : 1.1, // default : 1.0. C in SVM. 
    tol : 1e-5, // default : 1e-4. Higher tolerance --> Higher precision 
    max_passes : 20, // default : 20. Higher max_passes --> Higher precision 
    alpha_tol : 1e-5, // default : 1e-5. Higher alpha_tolerance --> Higher precision 

    kernel : { type: "polynomial", c: 1, d: 5}
    // default : {type : "gaussian", sigma : 1.0} 
    // {type : "gaussian", sigma : 0.5} 
    // {type : "linear"} // x*y 
    // {type : "polynomial", c : 1, d : 8} // (x*y + c)^d 
    // Or you can use your own kernel. 
    // kernel : function(vecx,vecy) { return dot(vecx,vecy);} 
});

the parameter C tells the SVM optimization how much you want to avoid misclassifying each training example.

I do not understand the other parameters.

解决方案

Just take a look at the equation of the soft-margin C-SVM:

It points out that C defines the trade-off between missclassifications and margin. This must be choosen sufficiently large depending on your data. What you'll also see here is the eps>0 parameter. This could possibly be your tolerance parameter and defines the error to the which is weighted by C parameter in the objective function.

For the kernel parameters, take a look at the dual problem for the SVM:

You'll see the term K(x_i,x_j). This is called the Kernel-Function. This function allows the SVM to learn non-linear descision boundaries. So if your data is not linearly separatable, you can use such a function to tranform your data, actually it's dot-product, into an higher dimensional feature space to separate them there. Just take a look at this guide, it will teach you the basics about the training process of an SVM and some best practices:

https://www.csie.ntu.edu.tw/~cjlin/papers/guide/guide.pdf

这篇关于训练SVM的参数是什么的文章就介绍到这了，希望我们推荐的答案对大家有所帮助，也希望大家多多支持IT屋！