Adaboost-DSNN: an adaptive boosting algorithm based on deep self normalized neural network for pulsar identification

Abstract

ABSTRACT A modern pulsar survey generates a large number of pulsar candidates. Filtering these pulsar candidates in a large astronomical data set is an important step towards discovering new pulsars. In this paper, a novel adaptive boosting algorithm based on deep self normalized neural network (Adaboost-DSNN) is proposed to accurately classify pulsar and non-pulsar signals. To train the proposed method on a highly imbalanced data set, the Synthetic Minority Oversampling TEchnique (SMOTE) was initially employed for balancing the data set. Then, a deep ensemble network combined with a deep self-normalized neural network and adaptive boosting was developed to train and learn the processed pulsar data. The design of the proposed Adaboost-DSNN method significantly reduced the computational time when dealing with large astronomical data sets, while also improving the classification performance. The scaled exponential liner units activation function was used to normalize the data. Considering their neighbour information and the special dropout technique (α-dropout), Adaboost-DSNN displayed good pulsar classification performance, while preserving the data properties across subsequent layers. The proposed Adaboost-DSNN method was tested on the High Time Resolution Universe Survey data sets (HTRU-1 and HTRU-2). According to experimental results, Adaboost-DSNN outperform other state-of-the-art methods with respect to training time and F1-score. The training time of the Adaboost-DSNN model is 10x times faster compared to other models of this kind.