With the development of intelligent technology, data in practical applications show exponential growth in quantity and scale. Extracting the most distinguished attributes from complex datasets becomes a crucial problem. The existing attribute reduction approaches focus on the correlation between attributes and labels without considering the redundancy. To address the above problem, we propose an ensemble approach based on an incremental information level and improved evidence theory for attribute reduction (IILE). Firstly, the incremental information level reduction measure comprehensively assesses attributes based on reduction capability and redundancy level. Then, an improved evidence theory and approximate reduction methods are employed to fuse multiple reduction results, thereby obtaining an approximately globally optimal and a most representative subset of attributes. Eventually, using different metrics, experimental comparisons are performed on eight datasets to confirm that our proposal achieved better than other methods. The results show that our proposal can obtain more relevant attribute sets by using the incremental information level and improved evidence theory.