来自线性 SVM 模型的权重(在 R 中)? [英] Weights from linear SVM model (in R)?

问题描述

使用 kernlab 我用如下代码训练了一个模型:

Using kernlab I've trained a model with code like the following:

my.model <- ksvm(result ~ f1+f2+f3, data=gold, kernel="vanilladot")

由于它是一个线性模型，我更喜欢在运行时将分数计算为特征值的简单加权总和，而不是使用完整的 SVM 机制.我怎样才能将模型转换成这样的东西(这里有一些虚构的权重):

Since it's a linear model, I prefer at run-time to compute the scores as a simple weighted sum of the feature values rather than using the full SVM machinery. How can I convert the model to something like this (some made-up weights here):

> c(.bias=-2.7, f1=0.35, f2=-0.24, f3=2.31)
.bias    f1    f2    f3 
-2.70  0.35 -0.24  2.31 

其中 .bias 是偏差项，其余的是特征权重?

where .bias is the bias term and the rest are feature weights?

这是一些示例数据.

gold <- structure(list(result = c(-1, -1, -1, -1, -1, -1, -1, -1, -1, 
-1, -1, -1, -1, -1, -1, -1, -1, -1, -1, -1, 1, 1, 1, 1, 1, 1, 
1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1), f1 = c(0, 0, 0, 0, 
0, 0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 0, 0, 0, 0, 0, 0, 1, 1, 0, 0, 
1, 0, 0, 0, 0, 0, 0, 1, 0, 1, 1, 0, 0, 0, 1), f2 = c(13.4138113499447, 
13.2216999857095, 12.964145772169, 13.1975227965938, 13.1031520152764, 
13.59351759447, 13.1031520152764, 13.2700658838026, 12.964145772169, 
13.1975227965938, 12.964145772169, 13.59351759447, 13.59351759447, 
13.0897162110721, 13.364151238365, 12.9483051847806, 12.964145772169, 
12.964145772169, 12.964145772169, 12.9483051847806, 13.0937231331592, 
13.5362700880482, 13.3654209223623, 13.4356400945176, 13.59351759447, 
13.2659406408724, 13.4228886221088, 13.5103065354936, 13.5642812689161, 
13.3224757352068, 13.1779418771704, 13.5601730479315, 13.5457299603578, 
13.3729010596517, 13.4823595997866, 13.0965264603473, 13.2710281801434, 
13.4489887206797, 13.5132372154748, 13.5196188787197), f3 = c(0, 
1, 0, 0, 0, 1, 0, 0, 0, 0, 0, 1, 1, 0, 0, 0, 0, 0, 0, 0, 1, 0, 
0, 0, 1, 0, 1, 1, 0, 0, 0, 0, 1, 0, 0, 0, 0, 1, 0, 0)), .Names = c("result", 
"f1", "f2", "f3"), class = "data.frame", row.names = c(NA, 40L
))

推荐答案

要获得偏差，只需使用全为零的特征向量评估模型.要获得第一个特征的系数，请使用第一个位置为1"且其他位置为零的特征向量评估模型 - 然后减去您已经知道的偏差.恐怕我不知道 R 语法，但从概念上讲，您想要这样的东西:

To get the bias, just evaluate the model with a feature vector of all zeros. To get the coefficient of the first feature, evaluate the model with a feature vector with a "1" in the first position, and zeros everywhere else - and then subtract the bias, which you already know. I'm afraid I don't know R syntax, but conceptually you want something like this:

bias = my.model.eval([0, 0, 0])
f1 = my.model.eval([1, 0, 0]) - bias
f2 = my.model.eval([0, 1, 0]) - bias
f3 = my.model.eval([0, 0, 1]) - bias

要测试您是否正确执行此操作，您可以尝试以下操作:

To test that you did it correctly, you can try something like this:

assert(bias + f1 + f2 + f3 == my.model.eval([1, 1, 1]))

这篇关于来自线性 SVM 模型的权重(在 R 中)?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！