Abstract:

Preference learning is one of the important methods in machine learning and has been applied in various fields. Preference learning in the presence of uncertain information is still a field that deserves further attention. The main objective of this paper is to construct a preference learning model for three‑parameter interval grey numbers (TPIGNs) using the UTilités Additives (UTA) method, to solve two key issues in preference learning: (1) failing to find the optimal preference, and (2) having multiple optimal preferences complicates the selection process. To address the two issues, this paper proposes a preference-learning method based on the UTA model for TPIGNs considering the inconsistent decision maker's pairwise preferences. TPIGNs are first converted into whitening values under a triangular possibility distribution and then incorporated into an additive value-function model. The first-rank acceptability index from stochastic multicriteria acceptability analysis (SMAA) is used to measure the prior importance of each reference alternative, from which an importance index is defined for each pairwise preference constraint. An improved UTA model is then constructed to identify and remove low-importance contradictory preferences through binary variables while retaining more informative preference statements. A random-direction extreme-point sampling procedure is used to obtain representative value-function parameters and improve the robustness of the ranking result. A supplier-selection case involving age-friendly equipment illustrates the implementation of the method and its advantage over the traditional UTA method in handling inconsistent preference information.