Since the distinguished style information of data may latently or obviously present in each data class in a given real-world dataset, a double knowledge representations based classification method (DKR-CM) from the perspective of social networks is proposed. In the training stage, a social network corresponding to all data samples in a dataset is easily built using the on-hand KNN method. In addition, style information of each data sample and each data class are respectively exploited in the social network. In the prediction stage, the proposed double knowledge representations (DKR) is utilized to improve the classification hebaviors of the DKR-SCM. In other words, each data sample is classified into the data class which it approaches to as far as possible from the perspectives of both physical features and style information of data. Experimental results demonstrate that the DKR-CM is at least comparative to the compared classification methods on the datasets with no or inapparent style information and outperforms them on the datasets with obvious style information.