It is a new approach for attribute reduction in the neighborhood decision system from the viewpoint of information theory. By analyzing the decision attribute values of samples in neighborhood, the inconsistent neighborhood matrix is defined. The inconsistent neighborhood matrix can be used to narrow the research range while adding more attributes with respect to existing condition attributes. Therefore, it is faster to calculate the significance of attributes by means of condition entropy. The relationship between the conditional entropy and the positive region is found under the neighborhood systems. An attribute reduction algorithm based on inconsistent neighborhood matrix is proposed, and the internal connection between this algorithm and other algorithms is analyzed. The experiment verifies the effectiveness of the proposed algorithm.