Today, people devote more of their time on social networks. In these fields, users need to make sure of activity together and ride them as a group called group recommendation systems. The primary objective of this approach is to propose one or more entities to a group of individuals to maximize the requests and benefits of that group of individuals. Collaborative filtering approaches are widely employed in these procedures and are based on a complete initial ranking in the user-item matrix. However, in the real system, this matrix is still sparse, and the priority of users is unknown. This problem can make memory-based collaborative filtering unsuitable for group recommendation systems. Many types of research have been done to solve these systems' cold start and sparsity problems. However, unlike the developed approaches that emphasize the problem of the sparse item-user matrix in individual recommendation systems, the approach of this research is solving this problem is the group recommendation systems and tries to provide an optimal solution for the sparse matrix of the user-item. The central part of the proposed method is based on a multilayer perceptron that computes the similarity between items. It is indicated that the proposed method gives group members more satisfaction with the other five existing algorithms.