Preference Inference: Reasoning About User Preferences in a Combinatorial Setting
In Computer-Aided Decision Making, it helps if the computer knows something about what options the user prefers. This includes, for example, recommender systems (which option should be shown to the user next?), and applications of multi-criteria decision making. However, preference elicitation is problematic: the user probably won't be happy to answer lots of questions just to elicit their preferences. Instead the system can learn partial information about their preferences from the choices they make; I call this process ‘preference inference’. This can be contrasted with preference learning, in which a user preference relation is learned from the input preferences.
Preference inference can vary considerably depending on the assumptions made on the user preference relation, and also on the input preference languages. I will discuss cases where the user models are weighted sums, and where they involve qualitative preference relations, especially those based on different forms of lexicographic order; I also discuss various kinds of input preference statement.
I gave a version of this talk as an invited tutorial at the SUM’17 conference (Scalable Uncertainty Management). It includes work done in collaboration with Insight PhD students Mojtaba Montazery and Anne-Marie George.
Wednesday, 6 December, 2017 - 15:00 to 16:00