With their rapid development, the scale and complexity of software systems are rapidly growing. Identifying and organizing files of similar functionality into the same module, which is called architecture recovery, contributes to the maintainability of a software system. Manual architecture recovery on large-sized software requires unbearable costs, hence a lot of automatic algorithms have been proposed in recent years. However, the accuracy of current algorithms is still not sufficient to support practical applications. To improve the accuracy of architecture recovery, this work proposes a novel algorithm NEGAR, which leverages random walks to extract latent graphic information from the dependency graph of files in the software system and learn the node representation for clustering. The proposed algorithm NEGAR has been comprehensively evaluated on three medium-sized and two large-sized and super large-sized software systems in terms of four widely-used metrics. The experimental results demonstrate the outstanding accuracy and excellent scalability of NEGAR.