A Variable Neighborhood Hybridizing for Better Ranking Models in Information Retrieval

Abstract

Abstract Variable Neighborhood Search (VNS) optimizes heuristic solutions for daily problems by adjusting neighboring solutions' systematic changes. This study uses two adaptations of VNS, utilizing four random probability distributions and tradition random number generator to fine-tune offspring solutions. The first generates solutions using a singular probability distribution, while the second perturbs each solution by randomly selecting a probability distribution for each mutation step size. The paper uses the two variants with the traditional variable neighborhood, combining variable neighborhood with gradient ascent and combining variable neighborhood with simulated annealing algorithms. From the obtained results, we can conclude that the second randomization method applied on combining variable neighborhood with a simulated annealing (VNARandom) method outperformed the other methods and recent research. In addition, this second randomization technique has a significant positive impact on the VNS methods for better evolving ranking models. Specifically, it has an improved impact on the recent GVN (combining Variable neighborhood with Gradient Ascent) method from the previous research studies. Although, it has a negative effectiveness impact on traditional VNS. This paper proposes novel VNS approaches as a contribution as well.