In 840 processor 590 optionally computes a mode-collapse reconstruction error for a first group of training features, for example group of training features 512, optionally by executing software object 531. Optionally, a target intersection group of features associated with set-operator prediction model 510A is computed by applying an intersection operator to at least two first groups of features, for example group of training features 512 and group of training features 513. Optionally, a target subtraction group of features associated with set-operator prediction model 510C is computed by applying a subtraction operator to at least two second groups of features, for example group of training features 512 and group of training features 513. Optionally, a target union group of features associated with set-operator prediction model 510B is computed by applying a union operator to at least two third groups of features, for example group of training features 512 and group of training features 513. Optionally, in 840 processor 590 provides group of training features 512 and group of training features 513 to set-operator prediction model 510A to produce an intersection group of features. Optionally, processor 590 provides group of training features 512 and group of training features 513 to prediction model 510C to produce a subtraction group of features. Optionally, processor 590 provides the subtraction group of features and the intersection group of features to prediction model 510B to produce a union group of features. Optionally, processor 590 applies a mean square error method to the union group of features and group of training features 512 to produce a mode-collapse reconstruction error score. Optionally processor 590 uses the mode-collapse reconstruction error score when computing the loss score, for example by adding the mode-collapse reconstruction error score to the sum of the plurality of model loss scores.