Attribute reduction is an important application in rough set theory, and it has been widely used in such areas as machine learning and data mining so far. Neighborhood rough set is a vital method for processing continuous data in rough set theory. For the existed detects of attribute reduction in the current neighborhood rough set model, the model of neighborhood rough entropy based on a neighborhood rough set is defined, meanwhile, the concepts of neighborhood rough combination entropy, neighborhood rough conditional entropy and neighborhood rough mutual information entropy are given, where the neighborhood rough mutual information entropy is an important method for evaluating the correlation of attribute sets, and at the same time, the neighborhood rough mutual information entropy is also proved to has a property of non-monotonic changing, therefore a non-monotonic attribute reduction algorithm based on neighborhood rough mutual information entropy is proposed. The experimental analysis show that the proposed algorithm has not only better results but also higher reduction efficiency than existing monotonic algorithms in attribute reduction.