Jesse Read, Ecole Polytechnique
In multi-label learning it is now widely known that we can often benefit by modelling labels together. In fact, this is arguably the raison être of multi-label learning research; to design, study, and apply such models (rather than independent models). The benefit can be in terms of more efficient computational performance, better interpretation/explainability, and/or better predictive performance; and the multi-label literature has been furnished with excellent examples of all of these backed by convincing experimental studies. But why do we get such a benefit? Answers are often phrased involving some variation of the term "label dependence", however there is a much more interesting story to tell. For example, independent models can often outperform dependence-based models, and on other occasions joint modelling of labels can provide a boost in accuracy even when labels are evaluated separately (and thus joint modelling should not be needed). In this talk we take a fresh look and further explore the question, with a specific emphasis on connecting to neighbouring areas, such as transfer and multi-task learning. In exploring what it means for one label or task to be 'dependent' on another and in which contexts, we take a path through some old and some new areas of the literature and, building from multi-label learning methodologies, we challenge (or in the least, put into question) some common assumptions made regarding label/task similarity.