To address the recommendation problems from malicious entities, a role separation based trust evaluation model (RSTrust) is proposed in this paper. In RSTrust, roles which entities act during trust evaluation are classified into transaction roles and recommendation roles. Trust on entity is therefore described as transaction trust and recommendation trust according to their associated roles, which leads to the separation of interference between different roles on different trusts. During the calculation of the global trust for an entity, the global recommendation trust of a recommender is used as a trust weight in RSTrust, and the disturbance of recommendation from malicious entities on global trust can be eliminated effectively. Analysis and simulation results demonstrate that RSTrust model has the fine feature of anti-malicious recommendation and good astringency.