CHARACTERIZING A CLASSIFICATION DIFFICULTY ASSOCIATED WITH A TARGET DATA ITEM

A dataset comprising a plurality of data items is received, wherein at least a portion of the plurality of data items is each associated with a corresponding class label from labels of a plurality of classes. For each class of the plurality of classes, a separate class-conditional reconstructor is trained on one or more of the data items associated with that class. For a target data item in the dataset having a target class label among the labels of the plurality of classes, a first reconstruction error is calculated using the class-conditional reconstructor trained for the target class label, a second reconstruction error is calculated using a class-conditional reconstructor trained for a class other than the target class label, a ratio of the first reconstruction error to the second reconstruction error is determined, and using the ratio, a classification associated with the target data item is characterized.