#multi-label classification

So, officially, I work on machine learning problems these days. The money for me is spotty, but I have always wanted to spend more time in what passes for AI work these days and this is an opportunity. Specifically I am working on the multi-label classification problem: given a set of instances (things to classify) and a set of labels (classifications) we try to assign labels to instances. For some simple cases this has been worked out in great detail. For instance, if there is only one label (A or not-A) this is binary classification, a well worked out problem with many methods of solution. If the classification is of instances to a fixed set of mutually exclusive (and exhaustive) labels it becomes multi-class classification, another area with lots of work already done. Multi-label classification (MLC) can be seen as the most general form of all of this. In MLC, one has an instance that must receive some number of labels, where the labels can co-occur (i.e. they are not mutually exclusive) and the exact number of labels per instance in not known a priori. So far my work has centered on known problem transformations--methods of converting multi-label problems into multi-class and binary problems. The mathematician's trick of reducing a new problem into the form of one he or she already knows how to solve. The two obvious tricks: treating each label as a separate problem (makes MLC into a bunch of binary problems) and treating each unique label combination as a separate class (MLC -> one multi-class problem). So far, these simple tricks work pretty well. We've established a baseline level of performance for our ongoing work. More as we go and as non-disclosure agreements allow.