On using the ColumnFeatures class in ecole.observation.NodeBipartiteObs

[Arun0013]

Please help me in understanding the use of ColumnFeatures class in ecole.observation.NodeBipartiteObs?

Is it to identify the features of each decision variable? If so, how do I use it in my code after collecting a NodeBipartiteObs object?

Example: Let, observation be a NodeBipartiteObs object returned by ecole environment. One way to get the column features is by using observation.column_features which returns a numpy array. However, this gives only the feature values and does not tell which value is for which feature. How do we know the feature that a value corresponds to? Do I somehow make use of ColumnFeatures for this prurpose?

[AntoinePrv]

Hello @Arun0013 ,

Yes, the ColumnFeatures is meant to be used the way you describe. It is an enumeration. If you never used enum before, you can think of it as a way to give a name to numbers.
So in this case, the number would be the axis1 index in the observation.column_features array, and the name would be the description of what that feature represent.

Example:
For observation, observation.column_features axis0 are the possible variables to branch on, and the axis1 some features of these variables.
The following gives us the name of all features and their index:

>> observation.column_features.__members__
{'objective': <ColumnFeatures.objective: 0>,
...
 'is_basis_zero': <ColumnFeatures.is_basis_zero: 18>}

Now to get the feature is_basis_zero, we can do

observation.column_features[:, int(observation.ColumnFeatures.is_basis_zero)]

This is sort of a cheap way of doing a pandas Dataframe.

[Arun0013]

Thanks, @AntoinePrv. This feature is super helpful. And, so is your comment. 👍