Abstarct: The increasing attractiveness and capabilities of cyberspace and the widespread use of personal electronic devices have made the Internet environment the mainstay backbone of many interactions between people and businesses. A great deal of interest in online social networks is a clear sample in this regard. The equivalent graph of these networks has a high volume of data due to a large number of nodes (humans) and edges (friendships and relationships between individuals or groups), so the processing of information on these large and complex networks requires efficient methods. Community detection, as one of the most essential branches of social network analysis, extracts and categorizes information of a network by identifying its constituent components. The applications of Community detection in the various sciences have led to the development of numerous methods with different aims and approaches. In the simplest case, the community detection - that is strongly related to the type of network and how to study it - can be considered as clustering or partitioning problem. However, with the increasing size and volume of information available in networks, membership of nodes in two or more communities seems inevitable. A set of algorithms that assign each node just to a single community ignores the overlap of nodes in them and eliminates an amount of information. In this thesis, we present a novel method for overlapping community detection. We have used an improved version of node representation and graph embedding, improved node Embedding (modeling input information to deep learning network), as well as using computational components of deep learning such as stacked Autoencoders. According to performance evaluation criteria in overlapping social networks, the proposed method works better than the present ones. Apart from this, the proposed method has a moderate time complexity compared to most existing methods, due to the use of deep learning technics, it can use powerful processors to accelerate its computations, maintains a large number of proximity measures, and ultimately reflects the local and global structure of graph nodes in detected communities.