Uncover: Toward Interpretable Models for Detecting New Star Cluster Members

Area of Science:

• Astronomy
• Machine Learning
• Data Science

Background:

• Identifying member stars in stellar clusters is crucial for astrophysical research.
• Manual methods for finding new members are time-consuming and inefficient.
• Novelty detection models offer potential but require careful hyper-parameter tuning.

Purpose of the Study:

• To present Uncover, an interactive tool designed to assist astronomers in identifying previously unknown member stars within stellar clusters.
• To provide a novel approach for selecting appropriate hyper-parameter sets for flexible novelty detection models.
• To empower astronomers to integrate their domain expertise into the machine learning model selection process.

Main Methods:

• Developed a five-step workflow to replace manual trial-and-error hyper-parameter optimization.
• Incorporated interpretable summary statistics that models must adhere to.
• Designed an interactive tool, Uncover, for data and task abstraction in astronomy.

Main Results:

• The proposed workflow effectively identifies suitable hyper-parameter sets for novelty detection models.
• Uncover facilitates the use of domain expertise in quantifying model performance.
• User studies with domain experts confirmed the tool's usability and usefulness.

Conclusions:

• Uncover offers a significant improvement over traditional methods for identifying member stars in stellar clusters.
• The tool promotes a shift towards astronomers actively engaging with and modifying machine learning models.
• This approach enhances the reliability and interpretability of machine learning applications in astronomical research.