Metaheuristics aim to balance solution quality and computational time when solving NP-hard problems. Many metaheuristics have been developed in the literature, such as Tabu Search, Simulated Annealing, and Variable Neighborhood Search (VNS). They generally share a common structure composed of three main procedures: initialization, diversification, and intensification. Most studies using metaheuristics pay little attention to the initialization phase, often relying on random sampling or basic heuristics as the initialization strategy. In this work, we focus specifically on the initialization phase. Given a set of candidate solutions, by analyzing the landscape around each solution according to certain criteria, our objective is to identify, by prediction, the solution in the most promising region of the landscape, which will be used to initialize our metaheuristic. One of the most studied and widely used metaheuristics is Variable Neighborhood Search (VNS). In this work, we employ VNS to solve the Job Shop Scheduling Problem (JSSP), using the Best Improvement descent method as the local search strategy. Our preliminary results have shown a good learning ability of the machine learning model (random forest), and the performance indicates that our approach is better in terms of the average gap to the optimum compared to the classical initialization strategy.