TSPRank: Bridging Pairwise and Listwise Methods with a Bilinear Travelling Salesman Model

TSPRank introduces a hybrid ranking method combining pairwise and listwise approaches, framing the problem as a Travelling Salesman Problem to enhance global information handling. It integrates as an additional component into existing models, outperforming both pure pairwise and listwise methods across diverse tasks like stock ranking, information retrieval, and historical events ordering. This approach leverages combinatorial optimization for improved ranking performance. TSPRank is the first LETOR model to frame ranking as a combinatorial optimization plus graph representation task, showing superior performance across various ranking tasks.