2020PASP..132i4502P

Pulsar detection using machine learning is a challenging problem as it involves extreme class imbalance and strong prioritization of high Recall. This paper is focused on automatic detection of astrophysical pulses in single-pulse searches, both within and across surveys. We use the output from the first stage of our previously developed two-stage Single-Pulse Event Group IDentification approach and focus on the second stage (i.e., classification of pulse candidates). Specifically, for the first time in time-domain single-pulse searches we (1) use boosting and deep learning algorithms for within-survey classification and (2) investigate cross-survey classification by using two transfer learning methods, trAdaBoost (instance-based) and fine-tuning (parameter-based). Our experimental results are based on two benchmark data sets, Green Bank Telescope Drift-scan (GBTDrift) and Pulsar Arecibo L-band Feed Array (PALFA)-extended, created from the GBTDrift survey and the PALFA survey. The main findings include: (1) Due to the emphasis on high Recall, F₄ measure is more appropriate performance indicator than the balanced F₁ measure. (2) For the GBTDrift benchmark, AdaBoost outperformed the other models with 98.5% Recall and F₄ = 0.942. For the PALFA-extended benchmark, RandomForest performed the best, with 91.6% Recall and F₄ = 0.890. (3) Models performance degraded significantly when they were used for pulsar classification across surveys. (4) Transfer learning improved cross-survey classification significantly when training data in the target data set were limited. When the GBTDrift benchmark was used as the target data set, fine-tuned SPEGnet models had the highest F₄ measure, while trAdaBoost models had the highest F₄ measure when PALFA-extended benchmark was used as the target data set. (5) Pulse classification was affected not only by the between-class imbalance (i.e., pulsar versus non-pulsar), but also by the within-class imbalance, which was more prominent in the case of the PALFA-extended benchmark due to the lack of labeled low signal-to-noise ratio pulsars.