Many approaches to recommendation have been used to collect user preferences in order to improve their navigation and transformation rate. In some cases, we do not have explicit information about users or products. The reconstruction of a user or product profiles, based on indirect elements, therefore becomes necessary. Our objective is to exploit factors underlying user-product interactions to enrich our recommendation engine. We propose here a new hybrid approach - matrix factorization and convolutional neural network - that responds to the problem posed by an industrial company in the tire industry. Our algorithm was evaluated on a set of real data extracted from an online comparator. Our study shows that our model is more efficient than state-of-the-art models and it has allowed us to study the impact of the different types of information used on the results.