[Paper Review] Photometric light curves classification with machine learning

Tatiana Gabruseva, Sergey Zlobin+1 more|arXiv (Cornell University)|Sep 10, 2019

Astronomical Observations and Instrumentation11 references4 citations

TL;DR

This paper presents a machine learning approach for classifying photometric light curves using gradient boosting on handcrafted features, data augmentation, and feature selection. It achieved top-tier performance in the PLAsTiCC challenge, with 88–100% accuracy for most astronomical objects, except for supernova types, which showed lower accuracy (33–71%) due to high confusion among subtypes, particularly SN Iax (class 52).

ABSTRACT

The Large Synoptic Survey Telescope will complete its survey in 2022 and produce terabytes of imaging data each night. To work with this massive onset of data, automated algorithms to classify astronomical light curves are crucial. Here, we present a method for automated classification of photometric light curves for a range of astronomical objects. Our approach is based on the gradient boosting of decision trees, feature extraction and selection, and augmentation. The solution was developed in the context of The Photometric LSST Astronomical Time Series Classification Challenge (PLAsTiCC) and achieved one of the top results in the challenge.

Motivation & Objective

• To develop an automated, high-accuracy classification system for photometric light curves from the upcoming LSST survey.
• To address the challenge of classifying diverse astronomical transients and periodic variables from massive, unbalanced datasets.
• To create a robust, single-model solution that avoids complex ensembles, enabling real-world deployability.
• To improve classification performance through effective feature engineering, selection, and data augmentation.

Proposed method

• Employed LightGBM, a gradient-boosted decision tree algorithm, for classification using engineered features from light curves and metadata.
• Extracted 100+ statistical, time-domain, and photometric features including flux skewness, median absolute deviation, autocorrelation, and color indices (e.g., g−r, r−z).
• Applied data augmentation techniques to reduce overfitting and improve generalization on the highly unbalanced training set.
• Used permutation importance and Pearson correlation to rank and select the most informative features, reducing redundancy.
• Performed 5-fold cross-validation to evaluate model stability and prevent overfitting.
• Integrated passband-specific features (e.g., flux_i, gauss_s_i, fitted_g_r) and redshift metadata (e.g., hostgal-photoz) into the final feature set.

Experimental results

Research questions

• RQ1 Can a single, efficient machine learning model achieve high classification accuracy on photometric light curves without relying on complex ensembles or test-time augmentation?
• RQ2 Which combination of handcrafted features and data augmentation techniques yields the best performance for classifying diverse astronomical transients and periodic variables?
• RQ3 How does feature selection impact model accuracy and generalization, especially on highly unbalanced datasets?
• RQ4 What are the main sources of misclassification, and can they be mitigated through targeted feature engineering or model architecture adjustments?

Key findings

• The model achieved 88–100% classification accuracy for all astronomical object types except for supernova subtypes.
• Supernova classes (42, 52, 62, 67, 90) showed the lowest accuracy, ranging from 33% to 71%, with class 52 (SNIax) being the most confused.
• The confusion matrix revealed significant overlap between different supernova types, especially SNIax and other core-collapse SNe.
• Feature selection using permutation importance and correlation filtering reduced overfitting and improved model performance.
• Data augmentation helped reduce overfitting and improved generalization, particularly on rare classes.
• Autoencoders and multiple parametric curve fits (e.g., Bazin, Karpenka) did not improve performance over single-fit or handcrafted features.