How to process data when each event has different number of available actions in VW CB?

[ruotaozhang]

Hi,

I was looking at the VW CB example given here.

In the example data, each event (i.e., row) has the same set of available actions which is [1,2,3,4,5]. Each action has a feature called "topic", and the topic of each of the 5 actions is extracted from the original json log to form 5 columns. Then the 5 columns are featurized and combined into a single column named "feature".

My question is: what if my data have different sets of actions for each event (i.e., row) in the json log? For instance, row 1 has two available actions [1,2], row 2 has three available actions [2,3,4], and row 3 has 1 available action [1]. In this case, how do I extract the action features and form the final "feature" column for the CB estimator?

Thank you!

AB#1837274

[eisber]

Working on it offline. We are waiting for the next VW release to get the required changes into the Java binding.