Road to robust prediction of choices in deterministic MCDM

• Published in: European journal of operational research Vol. 259; no. 1; pp. 229 - 235
• Main Authors: Pajala, Tommi, Korhonen, Pekka, Wallenius, Jyrki
• Format: Journal Article
• Language: English
• Published: Elsevier B.V 16.05.2017
• Subjects: Bayesian model; Linear value function; Multi objective programming; Multiple criteria decision making; Preference prediction; Multi objective programming; Multiple criteria decision making; Preference prediction; Bayesian model; Linear value function
• ISSN: 0377-2217; 1872-6860
• DOI: 10.1016/j.ejor.2016.10.001

•Linear value function approximation was found suitable for preference prediction.•Estimated and Analytic Hierarchy Process weights were the best prediction methods.•Bayesian model showed a small effect of linear consistency and value difference. We compare five different prediction methods (linear estimated weights, AHP weights, equal weights, logistic regression, and a lexicographic method) in their success rate for predicting preferences in pairwise choices. Students were asked to make pairwise comparisons between student apartments on four criteria: size, rent, travel time to the university and travel time to a (hobby) location of their choice. First ten choices were used to set up the estimation model, and subsequent ten choices are used for prediction. We find that the linear estimation method has the highest prediction success rate. Furthermore, the probability of predicting a choice correctly differs only slightly (by 0.1) between linear consistent and inconsistent subjects, ie. subjects whose preferences were consistent or inconsistent with a linear value function. This shows that in the absence of other preference information, a linear value function is suitable for prediction purposes.