Comparative analysis of machine learning techniques for feature selection and classification of Fast Radio Bursts

TL;DR

This study compares various unsupervised machine learning pipelines to classify Fast Radio Bursts, identifying features and candidate repeaters, thereby advancing understanding of their nature and aiding future observational efforts.

Contribution

It introduces a comprehensive evaluation of hybrid dimensionality reduction and clustering methods for FRB classification, emphasizing feature relevance and candidate identification.

Findings

Derived features improve classification accuracy

Certain non-repeater FRBs cluster with repeaters

Optimal hyperparameters enhance clustering performance

Abstract

Fast Radio Bursts (FRBs) are millisecond-duration radio transients of extragalactic origin, exhibiting a wide range of physical and observational properties. Distinguishing between repeating and non-repeating FRBs remains a key challenge in understanding their nature. In this work, we apply unsupervised machine learning techniques to classify FRBs based on both primary observables from the CHIME catalog and physically motivated derived features. We evaluate three hybrid pipelines combining dimensionality reduction with clustering: Principal Component Analysis (PCA) + k-means, t-distributed Stochastic Neighbor Embedding (t-SNE) + Hierarchical Density-Based Spatial Clustering of Applications with Noise (HDBSCAN), and t-SNE + Spectral Clustering. To identify optimal hyperparameters, we implement a comprehensive grid search using a custom scoring function that prioritizes recall while penalizing excessive cluster fragmentation and noise. Feature relevance is assessed using principal component loadings, mutual information with the known repeater label, and permutation-based F\textsubscript{2} score sensitivity. Our results demonstrate that the derived features, including redshift, luminosity, and spectral properties, such as the spectral index and the spectral running, significantly enhance the classification performance. Finally, we identify a set of FRBs currently labeled as non-repeaters that consistently cluster with known repeaters across all methods, highlighting promising candidates for future follow-up observations and reinforcing the utility of unsupervised approaches in FRB population studies.