Abstract: Since the lack of consideration of user heterogeneity in the of continuous opinion dynamic model, considering the opinion interactive process in reality, this paper improves the bounded confidence model from the immobilization of confidence threshold and convergence coefficient, as well as the undirected and unweighted opinion interaction, and constructs a opinion learning model for the heterogeneous members based on trust level. From the bigotry degree, the influence degree and the influence scope, three kinds of special members are defined in the opinion evolution process. Through the computational experiments based on multi-agent, the influences of three kind of special member opinion on the evolution of network group opinion are studied from the perspective of fans, authoritative members and high in-degree members. It is found that special members' characteristics do not directly affect the formation of group opinion, special members' characteristics paly a positive or negative adjustment role in the relationship between initial opinion of special members and the final average opinion of the group members.