The R package MLR supports multi-label classification that maps a feature vector into a set of discrete labels Y_1, Y_2,...,Y_k. For example, Y_1, ..., might be categorical demographic traits such as age, income, gender, and multiple of these may apply to a given example in the training data. This is sometimes called multi-task learning, I believe.
Some regression tasks, such as canonical correlation analysis, have a similar flavor, in which our labels are continuous and vector-valued. What is the best way to represent such a task in MLR? I have managed shoe-horn canonical correlation analysis into regular regression task, but I am badly abusing the predict and performance methods (I want to return a vector-valued prediction that is compared to vector-valued underlying truth).
Another approach would be to "vectorize" the training data, so that a K-valued target has each training example appearing K times. This loses some nice structure to the problem, predictions and performance evaluation, however.
