Consensus clustering-based undersampling for improved classification of transient events in time-domain astronomy...

Area of Science:

• Astronomy
• Computer Science
• Data Science

Background:

• Modern astronomy projects generate vast amounts of data, necessitating efficient analysis techniques.
• Classifying potential transient events from time-domain surveys is crucial for discovering new astronomical phenomena.
• High volumes of data and imbalanced training sets pose significant challenges in astronomical data analysis.

Purpose of the Study:

• To address the issue of imbalanced training data in the classification of astronomical transient events.
• To develop a method that effectively filters out false positives from astronomical survey data.
• To improve the accuracy and efficiency of analyzing high-volume astronomical datasets.

Main Methods:

• Investigated the classification of potential transient events from time-domain astronomical surveys.
• Evaluated oversampling methods and classifiers, noting a tendency towards overfitting.
• Proposed and applied a novel consensus clustering approach for undersampling majority-class instances.

Main Results:

• Oversampling methods combined with classifiers showed initial improvement but led to overfitting.
• Consensus clustering effectively undersampled majority-class instances, mitigating overfitting.
• The proposed method strengthens existing approaches by guiding the selection of representative samples.

Conclusions:

• Consensus clustering offers a robust solution for handling imbalanced data in astronomical transient event classification.
• This approach effectively filters false positives, reducing the need for laborious manual assessment.
• The study advances astronomical data analysis by providing an accurate and efficient method for handling big data challenges.