Combining Variable Neighbourhood with Simulated Annealing for Learning to Rank Problem

Abstract Variable Neighbourhood Search (VNS) is a problem-solving technique that improves heuristic solutions. The solutions are built around incremental adjustments to neighboring solutions. Changes are done during the climbing phase to get local optimal solutions, followed by a stochastic phase to achieve global optimum solutions. A range of mutation step-sizes are used in the exploration and exploitation methods. The function's purpose is to choose the best offspring to pass on to the next developing generation. As a consequence, this paper changes the Offspring solution in the following iteration applying a version of VNS based on four random probability distributions. Once the cooling temperature is met, it is capable of accepting a bad solution. Variable Neighborhood Annealing is a novel approach in Learning to Rank (LTR) (VNA). Each Offspring ranking model solution is built from a single probability distribution during the mutation phase (all mutation step-sizes made by only one probability distribution for each Neighbourhood candidate). Based on the results, we may infer that the VNA approach outperformed contemporary research on Evolutionary and Machine Learning methodologies. In the studies, we used datasets from Yahoo, Microsoft Bing Search (MSLR-WEB10K), and LETOR 4 (MQ2008, MQ2007).