Bayesian Utility Elicitation

Preference elicitation is an important component of interactive decision support systems that aim to make personalized recommendations to users by actively querying their preferences.
Assuming that the user's preferences are modeled by an utility function (unknown to the system), the goal is to ask informative questions to the user in order to determine the best alternative (or an alternative "close to" optimum) with only few questions.

One prominent approach is to adopt Bayesian decision theory for dealing with an uncertain utility function [Chajewska et al, 2000; Viappiani and Boutilier, 2010]. In this setting, the system assumes a prior distribution over the parameters of the utility function (the prior may be "uniformed", i.e. uniform, or "informative" depending on the context). Whenever the user provides new preference information, the distribution is updated according to Bayes (the likelihood of a response is evaluated according to a probabilistic "response model"). In order to evaluate which question is the most informative, we adopt the criterion of Expected Value of Information (EVOI). However how EVOI is compute depends on the exact type of questions (most previous work considered comparison queries); the computation of EVOI may be challenging in large datasets so approximation methods can be considered.

The internship will aim at improving the state of the art by:

- Considering a system that may pose different types of questions (ask to compare two alternatives, ask about preferences for specific attributes, ask to rank a number of items,..) dynamically choosing the most informative one, but also accounting for the "cognitive cost" of the question.

- Considering the recommendation of structured objects, such as providing a sequence of items in output. We want to adopt a principled semantics for rankings.

- Adapting the elicitation in context of multiple users (agents), in a way that the information acquired from one agent can be used to bootstrap the elicitation of another agent (in a setting similar to [Teso et al, 2017])

In addition to algorithmic development, it is expected that the student will perform experimentations to compare the efficiency of the proposed methods.

This internship would take place in the DECISION team of LIP6 and financed by ANR CoCoRICo-CoDec (Calcul, Communication, Rationalité et Incitations en Décision Collective et Coopérative).

REFERENCES:

* Stefano Teso, Andrea Passerini, Paolo Viappiani:
Constructive Preference Elicitation for Multiple Users with Setwise Max-margin. ADT 2017: 3-17

* Paolo Viappiani, Craig Boutilier:
Optimal Bayesian Recommendation Sets and Myopically Optimal Choice Query Sets. NIPS 2010: 2352-2360

* Shengbo Guo, Scott Sanner:
Real-time Multiattribute Bayesian Preference Elicitation with Pairwise Comparison Queries. AISTATS 2010: 289-296

*Urszula Chajewska, Daphne Koller, Ronald Parr:
Making Rational Decisions Using Adaptive Utility Elicitation. AAAI/IAAI 2000: 363-369