Feature Impact is computed for each column by creating new data with that column randomly permuted (but the others left unchanged), and seeing how the error metric score for the predictions is affected. The 'impactUnnormalized' is how much worse the error metric score is when making predictions on this modified data. The 'impactNormalized' is normalized so that the largest value is 1. In both cases, larger values indicate more important features. Elsewhere this technique is sometimes called 'Permutation Importance'.
GetFeatureImpact(model)
character. The model for which you want to compute Feature Impact, e.g.
from the list of models returned by ListModels(project)
.
Note that GetFeatureImpact
will block for the duration of feature impact calculation. If
you would prefer not to block the call, use RequestFeatureImpact
to generate an async
request for feature impact and then use GetFeatureImpactForModel
or
GetFeatureImpactForJobId
to get the feature impact when it has been calculated.
GetFeatureImpactForJobId
will also block until the request is complete, whereas
GetFeatureImpactForModel
will error if the job is not complete yet.