Interpretation of the base exit from Vowpal Wabbit

Question

I had a couple of questions about getting out of a simple VW launch. I read about the Internet and wikis, but I'm still not sure about a few basic things.

I ran the following Boston case data:

vw -d housing.vm --progress 1 

where the housing.vm file is configured as (partially):

and conclusion (partially):

Question 1:

1) Is it correct to consider the average loss column as the next step:

a) predict zero, so the first average loss is the squared error of the first example (with a prediction of zero)

b) build the model in example 1 and predict example 2. Average losses in 2 squares

c) build a model in example 1-2 and predict example 3. The average loss of 3 squares

d) ...

Do this until you click on the end of the data (assuming one pass)

2) What are the current function columns? This seems to be the number of non-zero features + interception. What is shown in the example assumes that the function is not considered if it is equal to zero - is that true? For example, the second entry has a value of 0 for "ZN". Does VW really look at this numerical function as missing?

Answer 1

Your statements are mostly true. By default, VW conducts online training, so in step c it takes the current model (weight) and updates it with the current example (instead of learning again in all the previous examples).

As you expected, the current function column is the number of (non-zero) functions for the current example. The interception function is automatically enabled if you do not specify --noconstant .

There is no difference between a missing function and a function with a zero value. Both that, and another means that you will not update the corresponding weight.