An unsuitable representation will make the task of mining classification rules very hard for a traditional evolutionary algorithm (EA). But for a given dataset, it is difficult to decide which one is the best representation used in the mining progress. In this paper, we analyses the effects of different representations for a traditional EA and proposed a growing evolutionary algorithm which was robust for mining classification rules in different datasets. Experiments showed that the proposed algorithm is effective in dealing with problems of deception, linkage, epistasis and multimodality in the mining task.