Using A One-Class SVM To Optimize Transit Detection

TL;DR

This paper demonstrates that One-Class SVMs, while slightly less accurate than CNNs, significantly outperform them in speed and resource efficiency for exoplanet transit detection, enabling faster analysis on standard hardware.

Contribution

The study introduces the use of One-Class SVMs for transit detection, showing they are faster and more resource-efficient than CNNs, with comparable accuracy.

Findings

One-Class SVMs are up to 84 times faster to fit than CNNs.

Prediction times for SVMs are over 3 times faster than CNNs.

SVMs can operate efficiently on non-GPU hardware.

Abstract

As machine learning algorithms become increasingly accessible, a growing number of organizations and researchers are using these technologies to automate the process of exoplanet detection. These mainly utilize Convolutional Neural Networks (CNNs) to detect periodic dips in lightcurve data. While having approximately 5% lower accuracy than CNNs, the results of this study show that One-Class Support Vector Machines (SVMs) can be fitted to data up to 84 times faster than simple CNNs and make predictions over 3 times faster on the same datasets using the same hardware. In addition, One-Class SVMs can be run smoothly on unspecialized hardware, removing the need for Graphics Processing Unit (GPU) usage. In cases where time and processing power are valuable resources, One-Class SVMs are able to minimize time spent on transit detection tasks while maximizing performance and efficiency.