Graph neural networks-based preference learning method for object ranking

• Published in: International journal of approximate reasoning Vol. 167; p. 109131
• Main Authors: Meng, Zhenhua, Lin, Rongheng, Wu, Budan
• Format: Journal Article
• Language: English
• Published: Elsevier Inc 01.04.2024
• Subjects: Graph neural networks; Learning to rank; Object ranking; Pairwise preference; Preference learning; Graph neural networks; Learning to rank; Preference learning; Object ranking; Pairwise preference
• ISSN: 0888-613X; 1873-4731
• DOI: 10.1016/j.ijar.2024.109131

Preference learning refers to the task of predicting the ranking of a collection of alternatives based on observed or revealed preference information. Object ranking is a critical problem within the domain of preference learning, which can be described as learning a ranking function based on training data in a ranked form. Some existing parametric preference learning methods are difficult to balance the relationship between expressive power and training cost. To deal with this issue, we introduce the concept of graph neural networks (GNNs) into preference learning, and propose a GNNs-based preference learning method. Our method is composed of three stages to cope with preference learning, i.e. preference relation graph construction, preference relation prediction, and object preference ranking. In the first stage, we map preference information onto the graph structure, and construct a directed graph with objects as nodes and preference relations between objects as edges. In the second stage, we formulate relation prediction as an edge classification problem on the graph, design a model consisting of multi-layer perceptron (MLP) and GNNs to extract edge features. In the third stage, we build a comparator neural network structure, which takes pairwise preference information as input and the score of each object as output, and the ranking of object scores is the objects' preference order. Experiments on preference learning datasets demonstrate that, compared to baselines, our method can achieve marked performance gains in terms of object ranking scenario.