svm - vowpal_wabbit 训练数据

Question

我试图了解训练和测试数据的 vowpal_wabbit 数据结构，但似乎无法理解它们。

我有一些训练数据，比如。

特征 1：0 特征 2：1 特征 3：10 特征 4：5 类别标签：A

特征 1：0 特征 2：2 特征 3：30 特征 4：8 类别标签：C

特征 1：2 特征 2：10 特征 3：9 特征 4：7 类别标签：B

我已经基于这个网站探索了一些训练数据示例。

http://hunch.net/~vw/validate.html

我的验证数据

1 | haha:1 hehe:2 hoho:3
1 | haha:2 hehe:2 hoho:3 
3 | haha:3 hehe:2 hoho:3 
1 | haha:4 hehe:2 hoho:3 
2 | haha:5 hehe:2 hoho:3  

但是，我不明白为什么它声称我分别有 4 个和 5 个功能。

验证：

验证反馈

Total of 5 examples pasted.

(example #1) Example “1 | haha:1 hehe:2 hoho:3”.
(example #1) Found “[label] |…” prefix format.
(example #1) Example label / response / class is “1”.
(example #1) Example has default “1.0” importance weight.
(example #1) Example has default “0” base.
(example #1, namespace #1) Using default namespace.
(example #1, namespace #1) Found 3 feature(s).
(example #1, namespace #1, feature #1) Label “haha”.
(example #1, namespace #1, feature #1) Value “1”.
(example #1, namespace #1, feature #2) Label “hehe”.
(example #1, namespace #1, feature #2) Value “2”.
(example #1, namespace #1, feature #3) Label “hoho”.
(example #1, namespace #1, feature #3) Value “3”.

(example #2) Example “1 | haha:2 hehe:2 hoho:3 ”.
(example #2) Found “[label] |…” prefix format.
(example #2) Example label / response / class is “1”.
(example #2) Example has default “1.0” importance weight.
(example #2) Example has default “0” base.
(example #2, namespace #1) Using default namespace.
(example #2, namespace #1) Found 4 feature(s).
(example #2, namespace #1, feature #1) Label “haha”.
(example #2, namespace #1, feature #1) Value “2”.
(example #2, namespace #1, feature #2) Label “hehe”.
(example #2, namespace #1, feature #2) Value “2”.
(example #2, namespace #1, feature #3) Label “hoho”.
(example #2, namespace #1, feature #3) Value “3”.
(example #2, namespace #1, feature #4) Label “”.
(example #2, namespace #1, feature #4) Using default value of “1” for feature.

(example #3) Example “3 | haha:3 hehe:2 hoho:3 ”.
(example #3) Found “[label] |…” prefix format.
(example #3) Example label / response / class is “3”.
(example #3) Example has default “1.0” importance weight.
(example #3) Example has default “0” base.
(example #3, namespace #1) Using default namespace.
(example #3, namespace #1) Found 4 feature(s).
(example #3, namespace #1, feature #1) Label “haha”.
(example #3, namespace #1, feature #1) Value “3”.
(example #3, namespace #1, feature #2) Label “hehe”.
(example #3, namespace #1, feature #2) Value “2”.
(example #3, namespace #1, feature #3) Label “hoho”.
(example #3, namespace #1, feature #3) Value “3”.
(example #3, namespace #1, feature #4) Label “”.
(example #3, namespace #1, feature #4) Using default value of “1” for feature.

(example #4) Example “1 | haha:4 hehe:2 hoho:3 ”.
(example #4) Found “[label] |…” prefix format.
(example #4) Example label / response / class is “1”.
(example #4) Example has default “1.0” importance weight.
(example #4) Example has default “0” base.
(example #4, namespace #1) Using default namespace.
(example #4, namespace #1) Found 4 feature(s).
(example #4, namespace #1, feature #1) Label “haha”.
(example #4, namespace #1, feature #1) Value “4”.
(example #4, namespace #1, feature #2) Label “hehe”.
(example #4, namespace #1, feature #2) Value “2”.
(example #4, namespace #1, feature #3) Label “hoho”.
(example #4, namespace #1, feature #3) Value “3”.
(example #4, namespace #1, feature #4) Label “”.
(example #4, namespace #1, feature #4) Using default value of “1” for feature.

(example #5) Example “2 | haha:5 hehe:2 hoho:3 ”.
(example #5) Found “[label] |…” prefix format.
(example #5) Example label / response / class is “2”.
(example #5) Example has default “1.0” importance weight.
(example #5) Example has default “0” base.
(example #5, namespace #1) Using default namespace.
(example #5, namespace #1) Found 5 feature(s).
(example #5, namespace #1, feature #1) Label “haha”.
(example #5, namespace #1, feature #1) Value “5”.
(example #5, namespace #1, feature #2) Label “hehe”.
(example #5, namespace #1, feature #2) Value “2”.
(example #5, namespace #1, feature #3) Label “hoho”.
(example #5, namespace #1, feature #3) Value “3”.
(example #5, namespace #1, feature #4) Label “”.
(example #5, namespace #1, feature #4) Using default value of “1” for feature.
(example #5, namespace #1, feature #5) Label “”.
(example #5, namespace #1, feature #5) Using default value of “1” for feature.

Answer 1

为什么它声称我分别有 4 个和 5 个功能

行尾的额外空格符号被http://hunch.net/~vw/validate.html解释为额外功能。（是的，示例中的最后一行有两个额外的空格。）请注意， validate.html 报告了额外功能的空名称：

(example #4, namespace #1, feature #4) Label “”.

请注意，validate.html是在 JavaScript 中实现的，并且完全独立于 VW 本身的实现（在 C 中）。VW 忽略尾随空格。您可以使用以下方法对其进行测试：

$ vw -P 1 < sample.data
...   
average    since         example     example  current  current  current
loss       last          counter      weight    label  predict features
1.000000   1.000000          1      1.0     1.0000   0.0000        4
0.522042   0.044084          2      2.0     1.0000   0.7900        4
1.838150   4.470366          3      3.0     3.0000   0.8857        4
1.488676   0.440255          4      4.0     1.0000   1.6635        4
1.270585   0.398217          5      5.0     2.0000   1.3690        4

因此，所有五个示例都被报告为具有 4 个特征（请参见最后一列）。为什么是四个？自动添加了一个额外的常量（拦截）功能。如果你不想要它，你可以使用vw --noconstant.