As detailed in the project description paper https://dspace.mit.edu/handle/1721.1/127410, as for feature importance, if a column is divided in to multiple columns, then it's more helpful to define the importance of the original column as the sum of the importance of the derived columns, instead of the max.
However, in the current implementation, (i) as the final determination of the lineage is done by a simple "score > 0" argument, which method we use is actually irrelevant. (ii) I believe the current implementation uses neither method. It iterates over the derived columns, and overwrite score_original_col = score_derived_col every time. So the final score is the score of the last derived column.
As detailed in the project description paper https://dspace.mit.edu/handle/1721.1/127410, as for feature importance, if a column is divided in to multiple columns, then it's more helpful to define the importance of the original column as the sum of the importance of the derived columns, instead of the max.
However, in the current implementation, (i) as the final determination of the lineage is done by a simple "score > 0" argument, which method we use is actually irrelevant. (ii) I believe the current implementation uses neither method. It iterates over the derived columns, and overwrite
score_original_col = score_derived_colevery time. So the final score is the score of the last derived column.