Extreme learning machine auto-encoder (ELM-AE) combines extreme learning machine (ELM) technology with auto-encoder (AE), which can learn data features unsupervised and overcome the expensive time consumption of parameter iterative adjustment. However, ELM-AE, which aims to minimize reconstruction errors, cannot effectively use the data category information in classification problems, resulting in features with poor category separability. In view of this phenomenon, this paper proposes a data classification-oriented feature learning method of extreme learning machine auto-encoder with category information (CELM-AE), which limits the inter class dispersion and intra class similarity of the projected feature vector to the objective function of ELM-AE, and can obtain the optimal data representation with more class resolution through analytical algorithm. The classification experiments of 6 UCI data sets are carried out using the feature representation based on CELM-AE, ELM-AE and AE respectively. The results show that the classification accuracy and stability of the data features obtained by CELM-AE under the two classifiers (ELM/KNN) are better than ELM-AE and AE, and the time cost is very small, which shows the advantages of CELM-AE in extracting the separable feature representation of data.