Abstarct: These days, with the increasing complexity and increasing dimensions of social networks, data mining and machine learning methods are used to analyze them. Social networks are very popular due to communication and interaction among users. Also, due to the addition of users or communication between them, these networks are dynamic in nature. The dynamic nature of networks provides a way to predict communication between users, called lixnk prediction. Most networks use lixnk prediction methods to automatically suggest high-accuracy acquaintances. Also, in some criminal networks, if a significant portion of the lixnks are made using lixnk prediction methods, the topology does not change and real lixnks are revealed in the criminal networks, which allows the prediction of specific criminal actions. Therefore, by further reducing the false positive rate of lixnk prediction, the possibility of improving accuracy in this area increases. In recent years, networks baxsed on deep learning in forecasting have achieved advanced performance. Among the existing networks, automatic encoder has been very successful in this field. In the proposed method, we first map the graph to smaller dimensions using the node embedded while preserving the graph structure and present the nodes in the form of vectors of real numbers. These vectors are then applied as input to the deep learning model, and finally we present a lixnk prediction system with appropriate performance to solve problems in the field of lixnk prediction. The algorithm presented on the facebook dataset has been tested and achieved an accuracy of 0.98, which is better than the previous best method, GCN.