Cost Sensitive Self-normalized Deep Convolutional Neural Network for Pulsars Selection

Machine learning and deep learning algorithms have been used to classify pulsar and non-pulsar signals in large radio astronomical datasets. These existing machine learning and deep learning algorithms improve classification efficiency, but they have limitations when dealing with large amounts of astronomical data, such as class imbalance and the polarization of high recall and precision. In this paper, a new classification method for dealing with the class imbalance problem as well as the data training process is proposed. The proposed method is based on a cost-sensitive, Scaled Exponential Linear Unit (SELU) activation function and a convolutional neural network (CNN). The proposed method is entitled (CS-SNDCNN). A cost matrix was generated and then applied to each misclassification depending on class distribution. After that, these costs were then applied to the training process in order to improve the final classification accuracy. The proposed method was compared with the state-of-the-art baseline models, which revealed that our model performed much better than existing approaches, almost gaining 3% and 4% better F1-score and precision, respectively.