On April 10, 2017, Dr. Sébastien Destercke (Université de Technologie de Compiège) will give a talk about "Making cautious decisions in large or structured output spaces: some results and challenges" in the context of the SFB 901.
Abstract:
Imprecise probabilistic approaches, that works with convex sets of probabilities (or equivalent models), are an interesting extension of classical probabilities when one wants to make cautious predictions and inferences, i.e., predict a set of classes or alternatives rather than one when information is missing. While such predictions are relatively easy to obtain for usual classification problems with few alternatives, many questions and challenges arise when considering complex or structured output spaces involved for instance in multi-label classificationor preference learning, and that can contain billions of alternatives. Some of these questions are: What form should take the predicted set? what interesting properties of the precise probabilistic case can I expect to preserve? how can I compute efficiently the predicted set, or an approximation thereof? In this talk, I will formalise these questions more precisely and will bring some answers for specific problems such as ordinal regression, multi-label predictions or label ranking. I will finish with some other similar problems for which similar questions, that remain unanswered, arise.