Rough set is an important method for attribute reduction problem. The size and dependency of reduction, which correspond to efficiency and precision of reduction respectively, are two significant indicators to evaluate a reduction. However, most existing work about attribute reduction method is mainly focused on maximizing the classification power, without considering the reduction size. The current algorithm finds the dependency between decision variables and condition attributes, however the correlation among the condition attributes is neglected. As a result, the reduct may include redundant attributes. This paper emphasizes the significance and correlation of reduced size and considers reduction performance on both scale and effective information at the same time. A multi-objective ensemble particle swarm algorithm based on decomposition (MEPSO/D) is proposed. Firstly, adaptive weight vectors are used to improve the decomposition performance of multi-objective optimization problem. Then, dynamic selection strategy is applied to fully explore the advantages of the ensemble algorithm. Finally, the framework of the multi-objective attribute reduction algorithm is proposed. Experiments on different data sets from UCI show that our proposed method is feasible and suitable for attribute reduction.