537. Preference learning augmented by graph structure for multiple criteria sorting with varying interactions
Invited abstract in session MA-8: Advances in preference learning methods, stream Multiple Criteria Decision Aiding.
Monday, 8:30-10:00Room: Clarendon SR 2.08
Authors (first author is the speaker)
| 1. | Zhen Zhang
|
| Institute of Systems Engineering, Dalian University of Technology | |
| 2. | Yuan Gao
|
| Dalian University of Technology |
Abstract
Data-driven preference learning has emerged as a trend in supporting multiple criteria sorting problems, extending the traditional multiple criteria decision aiding methods to elicit complex preferences, such as capturing interaction effects between criteria. However, most existing studies fail to capture complex dependencies among criteria, especially when criteria interactions may vary with criteria performance. This study proposes a novel preference learning model augmented by graph structure, which enhances the value-driven threshold-based sorting procedure within a neural network framework, aiming to model varying criteria interactions. The preference model considers both the main effects of individual criteria and the interaction effects of pairwise criteria. Specifically, the additive value function is used to evaluate the main effects of individual criteria by aggregating marginal value functions obtained through piecewise linear approximation on individual criteria. Detection and attention mechanisms within the graph structure are further designed to detect and quantify varying criteria interactions that are beneficial for predicting sorting outcomes. To conduct the sorting procedure, the preference model and ordinal class thresholds are learned using a threshold-based loss function. Numerical experiments on real-world and synthetic data sets are conducted to validate the performance of the proposed preference learning model.
Keywords
- Decision Analysis
Status: accepted
Back to the list of papers