A Novel Method to Create Synthetic Samples with Autoencoder Multi-layer Extreme Learning Machine

The imbalanced classification is an important branch of supervised learning and plays the important roles in many application fields. Compared with the sophisticated improvements on classification algorithms, it is easier to obtain the good performance by synthesizing the minority class samples so that the classification algorithms can be trained based on the balanced data sets. In consideration of the strong representation ability of multi-layer extreme learningmachine (MLELM), this paper proposes a new method to create the synthetic minority class samples based on auto-encoder ML-ELM (simplified as AE-MLELM-SynMin). Firstly, an AE-MLELM is trained to obtain the deep feature encodings of original minority class samples. Secondly, the crossover and mutation operations are preformed on the original deep feature encodings and a number of new deep feature encodings are generated. Thirdly, the synthetic minority class samples are created by transforming the new deep feature encodings with AE-MLELM. Finally, the persuasive experiments are conducted to demonstrate the effectiveness of AE-MLELM-SynMin method. The experimental results show that our method can obtain the better imbalanced classification performance than SMOTE, BorderlineSMOTE, Random-SMOTE, and SMOTE-IPF methods.