The Capacitated Vehicle Routing Problem (CVRP) is a variant of routing problems widely studied in the literature. Despite its complexity (NP-complete), it is now possible to reach near-optimal solutions even for large instances. The purpose is to design routes of vehicles, starting and finishing at the same location, called a depot, to serve all the customers, while minimizing the total length of the fleet. Moreover, the sum of the demands of the customers on a route can not exceed the capacity of the vehicle. We can find in the literature many algorithms and heuristics to solve the CVRP. Since optimal solutions are computationally expensive to find, most of the recent works focus on approximation algorithms. Moreover, recent trends try to integrate learning mechanisms into metaheuristics to improve both the quality of the solutions obtained and the speed of the heuristic. Nevertheless, it is possible to enhance heuristics and metaheuristics by integrating specific knowledge related to the problem. In the case of the CVRP, Arnold in his thesis trained a neural network to predict if a given solution was near-optimal or not. Thanks to the knowledge generated, he found relevant metrics to characterize the badness of an edge in a solution. Following this work, we propose to predict edges that appear in good solutions with a neural network and then we use these predictions to improve the well-known Clarke & Wright heuristic.