The present invention, in some embodiments thereof, proposes producing a new multi-label group of features, without explicitly specifying which features to include in the new multi-label group of features, by manipulating two or more groups of features, for example two or more groups of features corresponding with two or more of an existing plurality of digital images. According to some embodiments of the present invention, the two or more groups of features are manipulated by one or more prediction models, such that applying one or more classification models to a new multi-label group of features produced by the one or more prediction models in response to the two or more groups of features produces an output set of labels that is similar, according to one or more similarity tests, to a target set of labels that can be computed by applying one or more set operators to other sets of labels associated with the two or more groups of features. An example of a similarity test is comparing a difference between the output set of labels and the target set of labels to one or more threshold difference values. For example, when the difference between the output set of labels and target set of labels is less than one or more threshold difference values the output set of labels is considered similar to the target set of labels. Optionally, a similarity test comprises comparing an absolute value of a difference to one or more threshold difference values.