Preference learning in multicriteria decision analysis for sorting problems: new methods and applications

Abstract

Multicriteria decision analysis consists of a set of principles and tools developed to assist in resolution of complex decision problems. One of the main features that distinguishes multicriteria decision methods from other methods in the fields of operational research and statistics is its high degree of incorporation of decision makers' preferences. Usually, preference information should be elicited by decision-makers themselves. However, depending on the decision-making problem, decision makers may not want to state their views or may not be able to set preferences for all model parameters. Hence, different approaches have been developed for indirect elicitation of preferences, such as preference disaggregation and preference learning. The disaggregation approach learns preferences from examples of decisions. Despite the existence of different methods based on preference disaggregation, there are few methods capable of modeling interaction between criteria or hierarchically structured criteria, two important characteristics in real-life decision-making problems. Preference learning learns preferences from large data sets and is considered an area of both machine learning and multicriteria decision analysis. Data-driven approaches to preference elicitation and modeling have become increasingly important because of the increasing availability of data sets and the proliferation of semi-automated computer interfaces. The proposal of this research project is to develop new methods for elicitation and learning preference through data, also able to model interaction between criteria and hierarchically structured criteria. This research project, therefore, is at the interface between multicriteria decision and machine learning. The proposed methodologies will be applied in the context of the development of new indices for measuring social vulnerability in Brazil.