Skip to Main content Skip to Navigation

Learning preferences with multiple-criteria models

Abstract : Multiple-criteria decision analysis (MCDA) aims at providing support in order to make a decision. MCDA methods allow to handle choice, ranking and sorting problems. These methods usually involve the elicitation of models. Eliciting the parameters of these models is not trivial. Indirect elicitation methods simplify this task by learning the parameters of the decision model from preference statements issued by the decision maker (DM) such as “alternative a is preferred to alternative b” or “alternative a should be classified in the best category”. The information provided by the decision maker are usually parsimonious. The MCDA model is learned through an interactive process between the DM and the decision analyst. The analyst helps the DM to modify and revise his/her statements if needed. The process ends once a model satisfying the preferences of the DM is found. Preference learning (PL) is a subfield of machine learning which focuses on the elicitation of preferences. Algorithms in this subfield are able to deal with large data sets and are validated withartificial and real data sets. Data sets used in PL are usually collected from different sources and aresubject to noise. Unlike in MCDA, there is little or no interaction with the user in PL. The input data set is considered as a noisy sample of a “ground truth”. Algorithms used in this field have strong statistical properties that allow them to filter noise in the data sets.In this thesis, we develop learning algorithms to infer the parameters of MCDA models. Precisely, we develop a metaheuristic designed for learning the parameters of a MCDA sorting model called majority rule sorting (MR-Sort) model. This metaheuristic is assessed with artificial and real data sets issued from the PL field. We use the algorithm to deal with a real application in the medical domain. Then we modify the metaheuristic to learn the parameters of a more expressive model called the non-compensatory sorting (NCS) model. After that, we develop a new type of veto rule for MR-Sort and NCS models which allows to take criteria coalitions into account. The last part of the thesis introduces semidefinite programming (SDP) in the context of multiple-criteria decision analysis. We use SDP to learn the parameters of an additive value function model.
Document type :
Complete list of metadatas

Cited literature [176 references]  Display  Hide  Download
Contributor : Abes Star :  Contact
Submitted on : Thursday, September 22, 2016 - 6:51:10 PM
Last modification on : Wednesday, October 14, 2020 - 3:40:40 AM


Version validated by the jury (STAR)


  • HAL Id : tel-01370555, version 1


Olivier Sobrie. Learning preferences with multiple-criteria models. Other. Université Paris Saclay (COmUE); Université de Mons, 2016. English. ⟨NNT : 2016SACLC057⟩. ⟨tel-01370555⟩



Record views


Files downloads