[Paper Review] Machine Learning in Astronomy: a practical overview

Dalya Baron|arXiv (Cornell University)|Apr 15, 2019

Gamma-ray bursts and supernovae18 references136 citations

TL;DR

A practical overview of supervised and unsupervised machine learning techniques applied to astronomical data, with emphasis on data challenges, evaluation, and implementing common algorithms, including probabilistic extensions and deep learning considerations.

ABSTRACT

Astronomy is experiencing a rapid growth in data size and complexity. This change fosters the development of data-driven science as a useful companion to the common model-driven data analysis paradigm, where astronomers develop automatic tools to mine datasets and extract novel information from them. In recent years, machine learning algorithms have become increasingly popular among astronomers, and are now used for a wide variety of tasks. In light of these developments, and the promise and challenges associated with them, the IAC Winter School 2018 focused on big data in Astronomy, with a particular emphasis on machine learning and deep learning techniques. This document summarizes the topics of supervised and unsupervised learning algorithms presented during the school, and provides practical information on the application of such tools to astronomical datasets. In this document I cover basic topics in supervised machine learning, including selection and preprocessing of the input dataset, evaluation methods, and three popular supervised learning algorithms, Support Vector Machines, Random Forests, and shallow Artificial Neural Networks. My main focus is on unsupervised machine learning algorithms, that are used to perform cluster analysis, dimensionality reduction, visualization, and outlier detection. Unsupervised learning algorithms are of particular importance to scientific research, since they can be used to extract new knowledge from existing datasets, and can facilitate new discoveries.

Motivation & Objective

• Motivate the use of machine learning as a data-driven counterpart to traditional model-based analysis in astronomy due to the growth of big, complex datasets.
• Provide practical guidance on applying supervised and unsupervised ML to astronomical datasets, including preprocessing, evaluation, and algorithm selection.
• Highlight popular algorithms (SVM, Random Forest, shallow neural networks) and unsupervised techniques for clustering, dimensionality reduction, and anomaly detection.
• Discuss data challenges in forthcoming surveys and how ML can help detect, characterize, and classify astronomical objects.

Proposed method

• Describe evaluation metrics and model validation schemes for supervised learning, including training/validation/test splits and cross-validation.
• Discuss input data handling: feature selection, normalization, scaling, and dealing with imbalanced datasets.
• Present and explain the core supervised algorithms: Support Vector Machines, Decision Trees, Random Forests, and shallow Artificial Neural Networks.
• Explain probabilistic handling of uncertainties in features and labels with Probabilistic Random Forest.
• Outline unsupervised learning topics (distance metrics, clustering, dimensionality reduction, anomaly detection) and their scientific relevance.
• Address practical usage considerations with shallow vs. deep models and feature extraction capabilities of convolutional architectures.

Experimental results

Research questions

• RQ1 How can supervised ML be effectively trained, validated, and tested on astronomical data?
• RQ2 What are the practical considerations for preprocessing, feature selection, and handling imbalanced data in astronomy?
• RQ3 How do common ML algorithms (SVM, Random Forest, shallow NN) perform on typical astronomical tasks and what are their limitations?
• RQ4 What advantages do unsupervised methods offer for discovering new knowledge in large astronomical datasets?
• RQ5 How can uncertainties in measurements and labels be incorporated into ML models for astronomy?

Key findings

• Probabilistic Random Forest improves classification accuracy by up to 10% with noisy features and up to 30% with noisy labels compared to traditional Random Forests.
• Probabilistic Random Forest naturally handles missing values and varying noise characteristics across training and test sets.
• Random Forests generalize better than a single Decision Tree due to aggregation across trees, but standard RFs do not natively account for feature/label uncertainties.
• SVM is simple and robust but sensitive to feature scaling and potentially affected by irrelevant features, thus feature selection is recommended.
• Ensemble and deep learning approaches can leverage raw data and reduce the need for extensive feature engineering in some contexts.
• The document emphasizes unsupervised learning as particularly important for extracting new knowledge and enabling discoveries from large datasets.