This study proposed to use representational learning algorithms to improve the identification of people with autism. They have a type of brain-based disorder that is born with social defects and repetitive behaviors.According to recent data from the Centers for Disease Control, one in 68 children in the United States who have autism. Brain images of people with autism from two databases, each from several sites and universities worldwide, known as ABIDE.Diagnosing people with autism is one of the most critical goals in cognitive science research. The topic of discussion in this field is studying people with autism disorders and the brain areas that cause these disorders. Cognitive science proposes to help diagnose the disease by examining brain areas and comparing these areas in healthy people to people with the disease.One of the methods that have recently been considered is the connectome matrix. This approach has led to the analysis of brain graphs and their comparison using complex networks. The connection matrix is usually analyzed using a complex network. Numerous papers examining the brain using sophisticated network metrics have attempted to improve the diagnosis of autism. In recent years, attempts have been made to study this disease using in-depth learning methods, which consider the connection matrix an image. The nature of this matrix is not an image but a graph that has complex grid metrics. The methods introduced in this study try to optimize the processing and reduce the time complexity in the deep learning network by using representational learning to increase the accuracy to help diagnose people with autism. Therefore, in this dissertation, we have tried to improve the diagnosis of autism by using representational learning methods to analyze complex networks with deep learning methods. In this study, we were able to increase the accuracy in fMRI, Embedding, Representation learning, Node2vec, struct2vec, DeepWalk