In this paper, we discuss the application of the data mining tools to identify typical features for highly cited papers (HCPs). By integrating papers' external features and quality features, the feature space used to model HCPs was established. Then, a series of predictor teams were extracted from the feature space with rough set reduction framework. Each predictor team was used to construct a base classifier. Then the five base classifiers with the highest classification performance and larger diversity on whole were selected to construct a multi-classifier system (MCS) for HCPs. The combination prediction model obtained better performance than models of a single predictor team. 11 typical prediction features for HCPs were extracted on the basis of the MCS. The findings show that both the papers' inner quality and external features, mainly represented as the reputation of the authors and journals, contribute to generation of HCPs in future.