Inconsistencies in ideal_labeler.py

[bngksgl]

Hi,

When i look at the def label(self, feature) function in ideal_labeler.py, i see that it receives the features from X_train and returns the label from y_train. It is doing this by finding records that have the same features passed from X_train, and it takes the first occurrence, and returns the first occurance records' label. This results with inconsistencies if there are multiple records that have the same entries. Wouldn't be better if you take the position of the queried id? Or what is the logic behind it?

def label(self, query_id): return self.y[query_id]

[yangarbiter]

The assumption of ideal_labeler class is that the same feature will have the same label, which is quite realistic in most cases. ideal_labeler is just mimicking the scenario where you have a feature vector and you want to acquire the label of that specific feature vector, so the input is feature vector instead of query_id If your scenario really need to index the labels, you can always override the current label method with what you just wrote.

[bngksgl]

Thank you for your quick response, i changed it in my code :)

[hsuantien]

In [https://github.com/ntucllab/libact/commit/c3babc89ee4cc7672539de46f5a9267d16dccc1b] we changed the behavior of ideal_labeler to randomly sample one label when there are multiple choices.