As the tourism market expands and tourist demands diversify, personalized tourism recommendation systems are essential for enhancing tourist experience and satisfaction. Traditional travel package recommendation methods often focus solely on recommending the most preferred options to tourists, thereby prioritizing recommendation precision over diversity. They also often neglect factors such as the stochastic nature of travel time caused by varying traffic conditions. This paper addresses these issues by proposing a two-stage personalized travel package recommendation framework, named TPRIVNS. In the first stage, we propose an improved variable neighborhood search algorithm (QT_VNS), based on quality score and topic diversity, to construct a candidate pool of travel packages. The algorithm introduces two penalty factors (one for attraction quality score and one for topic similarity), and defines five neighborhood operators for package adjustment and optimization, with the objective of balancing average attraction quality and intra-package diversity. In the second stage, we propose a stochastic optimization model considering inter-package diversity for travel package recommendation, which incorporates chance constraints to handle stochastic travel time and is solved using the Sample Average Approximation (SAA) method. Finally, we evaluate the proposed TPRIVNS framework using real tourism data, assessing the method’s precision, diversity, and time reliability.
