The inherent complexity of social networks in terms of topological properties requires sophisticated methodologies to detect communities or clusters. Community detection in social networks is essential for understanding organizational structures and patterns in complex interconnected systems. Traditional methods face challenges in handling the scale and complexity of modern social networks, such as local optima trapping and slow convergence. This paper proposes a hybrid method to improve the accuracy of community detection, leveraging stacked auto-encoder (SAE) for dimensionality reduction and the Shuffled Frog Leaping (SFLA) as memetic algorithm for enhanced optimization alongside k-means clustering. The proposed method constructs a hybrid similarity matrix combining structural information and community-related features, followed by SAE to reduce dimensionality and facilitate efficient processing of high-dimensional data. SFLA optimizes the k-means clustering process, introducing adaptability and diversity to exploration of the solution space. Experimental results indicated its superior performance in terms of normalized mutual information (NMI) and modularity compared to existing approaches.