A LSTM-cBiGAN based hybrid sampling method for time series customer classification

Accurate customer classification is an important prerequisite for successful customer relationship management. With the development of the Internet, companies have obtained more time series data on consumers. Time series customer classification can extract the potential characteristics of customers and classify them accurately. One of the thorny problems facing the customer classification problem is class imbalance. However, there are few existing studies that address the issue of class imbalance in time series classification. So we introduce Generative Adversarial Networks (GAN), a deep learning model with strong performance on general data, to time series classification. A novel LSTM-cBiGAN based hybrid sampling method was proposed. In the oversampling part we use LSTM-cBiGAN to generate the data. In the undersampling part we proposed Time Series Features based Nearest Neighbor (TSFNN) undersampling algorithm. Experimental results on real-world time series customer classification show that our proposed hybrid sampling method outperforms other benchmark sampling methods. It provides an important solution to the class imbalance problem in time series customer classification.