Classifying Radio-Loud and Radio-Quiet Quasars With Novel PCA Based Regression Classifier

Abstract

The problem of quasar classification comes in the class of highly imbalanced classification problems since Radio-loud (RL) quasars are rare and make up only about 10% of the quasar population. In this work, we use the Sloan Digital Sky Survey-DR3 dataset and introduce a PCA-based regression pipeline designed to maximize recall for rare classes in class-imbalanced astronomical data. We demonstrate an effective methodology to identify the key features of the dataset and apply Principal Component Analysis (PCA) for dimensionality reduction. For the PCA transformed SDSS-DR3 dataset, first two components account for the 97% of the observed variance. We perform classification of Radio-Loud (RL) and Radio-Quiet (RQ) quasars with Random Forest Classifier (RFC), novel PCA based balanced linear regression classifier (PBC), Random forest integrated with SMOTE classifier and XGBoost classifier with threshold tuning. RFC achieves an overall accuracy of 92% while PBC achieves an overall accuracy of 62%. XGBoost achieves an overall accuracy of 72% and SMOTE integrated RFC achieves an accuracy of 85%. Higher precision is obtained for RQ quasars in all classification methods. For the RL class, RFC achieves a recall of 0.04, XGBoost achieves a recall of 0.39, SMOTE integrated RFC achieves a recall of 0.25 and PBC achieves a recall of 0.52 attributed to the balanced logistic regression. RFC and PBC achieve F1 score of 0.08 and 0.19 respectively for RL while XGBoost achieves an improved F1 score of 0.22 but at the cost of reduced recall of the RL class. SMOTE integrated RFC achieves a better F1 score of 0.21 over RFC and PBC. Overall results of classifiers point to extreme class imbalance between RQ and RL classes in the data set.