Pocket switched networks (PSN) is a kind of delay tolerant network (DTN) which transmits messages based on node movement and opportunity encounter. The traditional similarity index based prediction methods are not suitable for the link prediction of PSN due to the frequent topology change and time-various link in PSN. According to the characteristics of PSN node behavior, this paper analyzes connection time, connection duration, and the law of node accessing areas. Node similarity is categorized into sociality behavior similarity and movement behavior similarity. After comparing AUC and Precision with different weight for sociality behavior similarity and movement behavior similarity, similarity index based on node behavior patterns (SNBP) is proposed. The experiment results on MIT Reality and Dartmouth Campus datasets show that comparison with CN, AA, and Katz similarity index, the proposed similarity index has better precision.