Javascript must be enabled for the correct page display

Unsupervised feature selection in astronomical surveys

Verdult, Sander (2020) Unsupervised feature selection in astronomical surveys. Bachelor's Thesis, Astronomy.

[img] Text
thesis - Finished.pdf

Download (2MB)
[img] Text
Restricted to Registered users only

Download (97kB)


With astronomical surveys continuing to increase in scope and ambition, more information is becoming readily available to the astronomy community. Not only do we have a greater number of objects to analyse at once, but we also have more features (parameters) available to us. However, scientific models trying to accommodate all these features at once become either overly complex or inaccurate due to overfitting. In this project, we will look into one of these astronomical surveys (GAMA) and create a catalogue of galaxy features consisting both of photometric and spectral information. After preselection using noise ratios, primary chi-square and the Extended Isolation Forest (EIF) anomaly detection algorithm, we will explain and then apply two different Unsupervised Feature Selection techniques (Principle Feature Analysis and a Hybrid algorithm) to this dataset in an attempt to define the most important features of galaxies. We shall examine the findings made, discussing lessons learned and possible steps to improve similar projects. Finally, we shall select a subset of both "best features" and a randomly selected subset, and compare results of a K-means clustering algorithm to get an indication of the viability of these techniques. In the end, we conclude that the used algorithms offer potential, but that they will require further work and modifications to provide compelling clustering results.

Item Type: Thesis (Bachelor's Thesis)
Supervisor name: Peletier, R.F. and Saifollahi, T.
Degree programme: Astronomy
Thesis type: Bachelor's Thesis
Language: English
Date Deposited: 02 Sep 2020 07:14
Last Modified: 02 Sep 2020 07:14

Actions (login required)

View Item View Item