Cluster ensemble becomes a research focus due to its success in privacy protection, distributing computing and reusing knowledge. Furthermore, the noise and isolation have little effect on the final result. Thereare two contributions in this paper. First, by regarding every base clustering as one attribute of the original data, it has found that the algorithm based on that is more extendable and flexible. Second, it designs a latent variable cluster ensemble (LVCE) model in this way and infers the algorithm of the model with Markov chain Monte Carlo (MCMC) approximation. At the end of the paper, the experimental results show that the MCMC algorithm of LVCE has a better result and can show the compactedness of data points clustering.