The basic idea of the Variable Neighborhood Search (VNS) algorithm is to systematically explore the neighborhood of current solution using a set of predefined neighborhood structures. Since different problem instances have different landscape and complexity, the choice of which neighborhood structure to be applied is a challenging task. Different neighborhood structures may lead to different solution space. Therefore, this work proposes a learning mechanism in a Variable Neighborhood Search (VNS), refer to hereafter as a Variable Neighborhood Guided Search (VNGS). Its effectiveness is illustrated by solving a course timetabling problems. The learning mechanism memorizes which neighborhood structure could effectively solve a specific soft constraint violations and used it to guide the selection of neighborhood structure to enhance the quality of a best solution. The performance of the VNGS is tested over Socha course timetabling dataset. Results demonstrate that the performance of the VNGS is comparable with the results of the other VNS variants and outperformed others in some instances. This demonstrates the effectiveness of applying a learning mechanism in a VNS algorithm.