Redshift‐Agnostic Machine Learning Classification: Unveiling Peak Performance in Galaxy, Star, and Quasar Classification (Using <scp>SDSS DR17</scp>)

ABSTRACT Classification of galaxies, stars, and quasars using spectral data is fundamental to astronomy, but often relies heavily on redshift. This study evaluates the performance of 10 machine learning algorithms on SDSS data to classify these objects, with a particular focus on scenarios where redshift information is unavailable. Leveraging features such as “z,” “u,” “g,” “r,” “i,” and redshift, we assess the accuracy of various algorithms, including XGBoost, Random Forest, and recurrent neural networks (RNNs). Our analysis demonstrates the superior accuracy of the Random Forest classifier when redshift is included. The feature importance analysis reveals that “redshift” is the most critical feature, contributing 64.7% to the classification accuracy, followed by the “z” band (10.0%) and the “g” band (7.95%). However, even in the absence of redshift, XGBoost, Random Forest, and RNNs exhibit promising results, indicating the potential of photometric data for accurate classification. We systematically compare classification outcomes with and without redshift, revealing the relative importance of different features and identifying the most robust classifiers for redshift‐limited scenarios. This research not only highlights the power of machine learning for astronomical classification but also provides a framework for reliable classification when redshift data is lacking. By uncovering the distinguishing spectral characteristics of galaxies, stars, and quasars that are independent of redshift, we open new avenues for efficient and accurate classification in large‐scale photometric surveys and the study of faint, high‐redshift objects.