Pulsar candidate selection using pseudo-nearest centroid neighbour classifier

Abstract

ABSTRACT A typical characteristic of the pulsar candidate classification task is the class imbalance between true pulsars and false candidates. This imbalance has negative effects on traditional classification methods. In this study, we introduce a strategy using a scatter matrix-based class separability measure to estimate the harmfulness of class imbalance on pulsar candidate classification. The measure quantitatively describes the damage of the imbalanced situations on the pulsar candidate classification problem and provides some priori information to guide us to select an appropriate data processing method and to construct an effective classifier. After that, we present a non-parametric data exploration technique, a pseudo-nearest centroid neighbour classifier (PNCN), to identify credible pulsar candidates from pulsar survey data sets. The PNCN algorithm can effectively resolve the class imbalance problem and is applicable to data streams. The proposed algorithm is tested on High Time Resolution Universe Pulsar Survey (HTRU) 2 (obtained by an analysis of HTRU Medium Latitude data) and LOTAAS 1 (obtained from the LOFAR Tied-Array All-Sky Survey). The experimental results show that the proposed classifier can excellently identify the pulsars with high performance: the precision and the recall on HTRU 2 are 92.3 per cent and 83.1 per cent, and those on LOTAAS 1 are 97.4 per cent and 95.6 per cent, respectively; the false positive rate (FPR) on HTRU 2 is 0.7 per cent, on LOTAAS 1 is 0.03 per cent, which is an order of magnitude lower than the corresponding FPR obtained in Lyon et al. (2016) and Tan et al. (2018).