Abstract: Social network analysis in Enron corpus found that the real e-mail network was a scale-free and small world in some degree. Then a spam collaborative filtering method was designed based on users' interaction. By adjusting the parameter λ, users can decide filtering spam by themselves or others or trade-off between them. Even in the absence of reading habits of users, the collaborative filtering method could achieve good performance. Because the Enron corpus was unlabeled, by adding i.i.d. assumption constraint to training data set W and test data set T, we labeled Enron corpus using improved EM (Expectation maximization) algorithm in a sense of minimum statistical risk in W ∪ T. Experiment results showed that the collaborative filtering method is simple and effective which can steadily increase average accuracy compared with single machine and ensemble filterings.