Efficient preference learning for AI-powered group decision-making

Abstract

Moreover, this dissertation proposes a group decision-making framework with consists of all the three aforementioned phases: a cost-effective preference elicitation method that actively collects informative preferences from people, a composite marginal likelihood method for preference learning, and group decision prediction using randomized voting rules. Simulations demonstrate the efficacy of this framework.

Much of this dissertation focuses on the efficient learning of ranking models. The main theoretical contributions are characterizations of (non)-identifiability of mixture models under different types of data, which relate to the explainability of the learned parameter and consistent algorithms. Guided by the identifiability results, generalized-method-of-moments-based algorithms and composite marginal likelihood methods are designed and compared.

A modern group decision-making paradigm involves three phases: data collection, preference learning, and preference aggregation. All three phases can be very challenging due to the large number of alternatives and resource constraints. How can we actively collect the most informative data with low costs? How can we learn from various structures of preferences efficiently and predict unknown preferences? When can we confidently explain the learning outcome? How can we make a group decision from uncertain predictions? This dissertation aims at answering these questions.