Searching for Needles in the Cosmic Haystack

Searching for pulsar signals in radio astronomy data sets is a difficult task. The data sets are extremely large, approaching the petabyte scale, and are growing larger as instruments become more advanced. Big Data brings with it big challenges. Processing the data to identify candidate pulsar signals is computationally expensive and must utilize parallelism to be scalable. Labeling benchmarks for supervised classification is costly. To compound the problem, pulsar signals are very rare, e.g., only 0.05% of the instances in one data set represent pulsars. Furthermore, there are many different approaches to candidate classification with no consensus on a best practice. This dissertation is focused on identifying and classifying radio pulsar candidates from single pulse searches. First, to identify and classify Dispersed Pulse Groups (DPGs), we developed a supervised machine learning approach that consists of RAPID (a novel peak identification algorithm), feature extraction, and supervised machine learning classification. We tested six algorithms for classification with four imbalance treatments. Results showed that classifiers with imbalance treatments had higher recall values. Overall, classifiers using multiclass RandomForests combined with Synthetic Majority Oversampling TEchnique (SMOTE) were the most efficient; they identified additional known pulsars not in the benchmark, with less false positives than other classifiers. Second, we developed a parallel single pulse identification method, D-RAPID, and introduced a novel automated multiclass labeling (ALM) technique that we combined with feature selection to improve execution performance. D-RAPID improved execution performance over RAPID by a factor of 5. We also showed that the combination of ALM and feature selection sped up the execution performance of RandomForest by 54% on average with less than a 2% average reduction in classification performance. Finally, we proposed CoDRIFt, a novel classification algorithm that is distributed for scalability and employs semi-supervised learning to leverage unlabeled data to inform classification. We evaluated and compared CoDRIFt to eleven other classifiers. The results showed that CoDRIFt excelled at classifying candidates in imbalanced benchmarks with a majority of non-pulsar signals (\u3e95%). Furthermore, CoDRIFt models created with very limited sets of labeled data (as few as 22 labeled minority class instances) were able to achieve high recall (mean = 0.98). In comparison to the other algorithms trained on similar sets, CoDRIFt outperformed them all, with recall 2.9% higher than the next best classifier and a 35% average improvement over all eleven classifiers. CoDRIFt is customizable for other problem domains with very large, imbalanced data sets, such as fraud detection and cyber attack detection