Evidence theory is regarded as an efficient tool to deal with uncertain or imprecise data and has been widely applied in many fields of information fusion. As an important part of evidence theory, Dempster's rule of combination is usually used to fuse different pieces of evidence so as to make a final decision. However, it cannot always acquire reasonable results. To resolve such a problem, a discounting approach based on cosine similarity is put forward. Each BBA (basic belief assignment) corresponding to piece of evidence is seen as a vector. The cosine of the angle between two vectors (cosine similarity) can effectively measure the similarity between two pieces of evidence. Thus, cosine similarity is applied to generate discounting factor. After discounting pre-processing, the discounted pieces of evidence are combined by Dempster's rule to obtain the final result. Example shows that the discounting approach presented has better performance than some other combination methods.