Combining Variable Neighbourhood with Gradient Ascent for Learning to Rank Problem
Osman Ali Sadek Ibrahim, Eman M. G. Younis · Research Square · 2022
Abstract Variable Neighbourhood Search (VNS) is used to optimize the solutions for heuristic problems. Solutions are based on neighboring solution systematic changes. The changes are made during the ascending phase to get local optimum and the perturbation phase to gain the global optimum solutions. Exploration and Exploitation procedures are made through various mutation step-size. The objective function to choose the best offspring to be move to the next evolving generation. Thus, this paper uses a variation of VNS based on four random probability distributions with Gradient Ascent for adapting the mutation of Neighborhood solution in the next iteration. It has the ability moderate the next proposed neighbourhood based of the Parent and Offspring fitness values. We called this novel approach in Learning to Rank (LTR) as Gradient Variable Neighbourhood (GVN). This variant generates each Offspring ranking model solution from one probability distribution in the whole mutation procedure (all mutation step-sizes made by only one probability distribution for each Neighbourhood candidate). From the obtained results, we can conclude that the GVN method outperformed the recent research of Evolutionary and Machine Learning methods. In the experiments, we used Yahoo, Microsoft Bing Search (MSLR-WEB10K) and LETOR 4 (MQ2008, MQ2007) datasets.