# Imbalance Learning for Variable Star Classification

**Authors:** Zafiirah Hosenie, Robert Lyon, Benjamin Stappers, Arrykrishna Mootoovaloo, Vanessa McBride  
**arXiv:** [2002.12386](https://arxiv.org/abs/2002.12386)  
**Date:** 2020-02-27  
**Summary page:** https://www.summarizepaper.com/en/arxiv-id/2002.12386v1/

## Key points

- Researchers address the challenge of accurately classifying variable stars into sub-types due to imbalanced learning.
- Previous research introduced a hierarchical machine learning classifier that showed promising results on CRTS data.
- The study aims to enhance hierarchical classification performance by incorporating data-level approaches for under-represented classes.
- Three data augmentation methods were experimented with: RASLE, GpFit, and SMOTE.
- Combining algorithm-level and data-level approaches led to a 1-4% improvement in variable star classification accuracy.
- Utilizing GpFit within the hierarchical model yielded a higher classification rate.
- Additional enhancements are needed for metric scores improvement, including a more robust standard set of identified variable stars and enhanced features.

## Summary

, , , , 

In their work on "Imbalance Learning for Variable Star Classification," Zafiirah Hosenie, Robert Lyon, Benjamin Stappers, Arrykrishna Mootoovaloo, and Vanessa McBride address the challenge of accurately classifying variable stars into their respective sub-types. They highlight the difficulty faced by machine learning solutions due to imbalanced learning, which leads to poor generalization performance, especially for rare variable star sub-types. In previous research, the authors developed a hierarchical machine learning classifier to overcome these deficiencies. This 'algorithm-level' approach showed promising results on Catalina Real-Time Survey (CRTS) data, surpassing traditional binary and multi-class classification schemes in this domain. Building upon their previous work, the researchers aim to enhance hierarchical classification performance by incorporating 'data-level' approaches to augment training data and better represent under-represented classes. They experiment with three data augmentation methods: Randomly Augmented Sampled Light curves from Magnitude Error (RASLE), light curve augmentation using Gaussian Process modeling (GpFit), and the Synthetic Minority Over-sampling Technique (SMOTE). By combining the 'algorithm-level' hierarchical scheme with the 'data-level' augmentation techniques, they achieve a further 1-4% improvement in variable star classification accuracy. The study reveals that utilizing GpFit within the hierarchical model yields a higher classification rate. However, the authors acknowledge that additional enhancements are required for metric scores improvement. They suggest the need for a more robust standard set of correctly identified variable stars and potentially enhanced features to continue advancing variable star classification accuracy. The paper is accepted for publication in Monthly Notices of the Royal Astronomical Society (MNRAS) and provides valuable insights into addressing imbalance learning challenges in astronomical data analysis.