using vw varinfo

Ariel Faigon edited this page Jul 24, 2018 · 16 revisions

Introduction

vw-varinfo is a small wrapper around vw which exposes all variables of a model in human readable form. The output includes the input variable names, including name-spaces where applicable, the vw hash value, the range [min, max] of the variable values in the training-set, the final model (regressor) weight, and the relative distance of each variable from the best constant prediction.

The wrapper is written in perl and can be found under utl/vw-varinfo in the source tree. vw-varinfo calls vw so vw should be installed somewhere in your PATH for vw-varinfo to work properly.

Here's a self-explanatory example output showing which foods affect increase vs decrease in daily weight during a diet:

FeatureName                HashVal   MinVal   MaxVal    Weight   RelScore
^bread                      220390     0.00     2.00   +0.0984     55.36%
^icecream-sandwich           39873     0.00     1.00   +0.0951     53.44%
^snapple                    129594     0.00     2.00   +0.0867     48.61%
^trailmix                   215350     0.00     1.00   +0.0708     39.53%
^peanut                     187714     0.00     8.00   +0.0464     25.49%
^pizza                       32162     0.00     4.00   +0.0420     23.00%
Constant                    116060     0.00     0.00   +0.0020      0.00%
^peas                       126345     0.00     1.00   +0.0002     -1.01%
^quinoa                     179283     0.00     1.00   +0.0002     -1.01%
^salmon                     215140     0.00     2.00   -0.1015    -59.42%
^chicken                    189058     0.00     1.00   -0.1133    -66.17%
^salad                      171971     0.00     2.00   -0.1256    -73.24%
^mayo                       187932     0.00     1.00   -0.1570    -91.28%
^oliveoil                    69559     0.00     1.00   -0.1570    -91.28%
^egg                           565     0.00     1.00   -0.1722   -100.00%

The example shows that based on the (simplistic) training set that was passed to vw-varinfo, eating egg and olive oil has the biggest negative correlation with weight-increase, while bread, icecream and sweetened drinks are the biggest enemies of weight loss. YMMV.

Usage:

vw-varinfo data.train

where data.train is a standard vw training-set. Just like vw itself, you may call vw-varinfo without any arguments to get a brief usage message.

more elaborate usage examples:

If you want to call vw with more arguments, simply pass them through to the training phase of vw like this:

vw-varinfo --l1 0.0005 -c --passes 40 data.train

Another example. Say you want to find the strength of certain interactions between two groups of features with respect to the output label. Assuming your data-set has two input-feature name-spaces starting with 'X' and 'Y', which separate your input features into two groups, you may run:

vw-varinfo -q XY your_data_set

And vw-varinfo will output all the pairs of interactions between features in name-space X and the features in name-space Y, ordered by their relative effect. You may add additional parameters to pass to vw training phase.

Sanity check of vw-varinfo using a contrived example

Here's a contrived example showing how vw-varinfo performs on a perfect linear model.

Step 1) we write a script generate-trainset.pl which loops 1000 times, in each loop iteration, it generates 5 random variables named 'a' through 'e' each with a random value in the interval [0 .. 1] and calculates the label y as:

y = a + 2*b + 3*c + 4*d + 5*e

Obviously, 'e' is 5 times more "important" than 'a' in affecting the value of y.

Here's the generate-trainset.pl script:

#!/usr/bin/perl -w
my $N = 1000;
for ($i = 1; $i <= $N; $i++) {
    my $a = rand(1);
    my $b = rand(1);
    my $c = rand(1);
    my $d = rand(1);
    my $e = rand(1);
    my $y = $a + 2*$b + 3*$c + 4*$d + 5*$e;
    printf "%g | a:%g b:%g c:%g d:%g e:%g\n", $y, $a, $b, $c, $d, $e;
}

Step 2) We run the script and save its output in a training-set:

$ generate-trainset.pl > abcde.train

Step 3) Running vw-varinfo on the training-set we get:

$ vw-varinfo  abcde.train
FeatureName        HashVal   MinVal   MaxVal    Weight   RelScore
^e                  180798     0.00     1.00   +5.0000    100.00%
^d                  193030     0.00     1.00   +4.0000     80.00%
^c                  140873     0.00     1.00   +3.0000     60.00%
^b                  244212     0.00     1.00   +2.0000     40.00%
^a                   24414     0.00     1.00   +1.0000     20.00%
Constant            116060     0.00     0.00   +0.0000      0.00%

which is exactly what was expected.

IOW: vowpal_wabbit perfectly figured out our formula, without knowing it in advance, by looking at the training data alone. QED.

Exercise:

  • modify the generate-trainset.pl so that y is always larger or smaller by some constant value.
  • how do you expect the result to be affected?
  • run your modified script, run vw-varinfo on the newly generated train-set and verify your hypothesis.

Credit & History:

vw-varinfo was written by Ariel Faigon, with help and guidance from John Langford. It was the main tool used for a little weight-loss project in its very early days.

Known issues:

vw-varinfo was written at about 2012, many options have been added to vw since. It is known not to work as-expected when some of the newer options are used, in particular, --cb and derivatives. Unfortunately, the code base is overly-complex and considered non-salvageable by its author.

There's a much cleaner, simpler and newer version of vw-varinfo in here, called vw-varinfo2. This version is a rewrite in python, is more efficient and more general. It is almost completely agnostic to vw options, so it should be easier to maintain going forward, but it doesn't yet support multi-class. If you want to enhance it, I (ariel) suggest that you start with this code-base.

You can’t perform that action at this time.
You signed in with another tab or window. Reload to refresh your session. You signed out in another tab or window. Reload to refresh your session.
Press h to open a hovercard with more details.