A large number of human behaviors occur on the Internet, which has become the most important virtual space corresponding to the real space. Social stratification research in the traditional virtual space is based on objective indicators such as the opportunity and ability of network information resources possession, but does not involves the specific behavior of users using network resources and the content and nature of information and other factors. This paper makes use of the big data of users’ online behaviors provided by China Internet Network Information Center, investigates and analyzes the characteristics and differences of users’ online behaviors of different levels in the virtual space from two aspects of online time and content. The study finds that different levels of users’ time spent in the virtual space and their attention focuses are very different. Higher level users can make better use of network resources to work and shopping, and stay in the virtual space has a relatively stable time. However, lower level users spend a lot of attentions on leisure and entertainment applications, and the stay time is not stable. In addition, this paper uses the neural network model (W2V-BP) based on Word2vec to conduct social stratification recognition of users' online behavior data in the virtual space, and the recognition accuracy reaches 90.22%, indicating that there are behavioral characteristics in the virtual space that can distinguish social stratification.