SDSS-DR14-Classification-of-Stars-Galaxies-and-Quasars

Classification of Stars, Galaxies and Quasars using Sloan Digital Sky Survey DR14

Project Description

The focus of this project is to compare the performance of a traditional machine learning (ML) technique with the performance of a modern neural network (NN) to solve a classification problem. Using a real-world dataset from SDSS, we explore how to build the different the models and analyse the factors that influence the success of training. As raw optical images of the stars, galaxies and quasars in the survey can appear visably similar as 'point sources' on the sky, it can prove diffcult to calssify the type of object being observed. Therefore, the testing and implementation of ML and NN techniques on a dataset like SDSS DR14 can a provide quick and effective method for object classification. The results from the ML and NN models can be compared to the class column in the dataset to evaulte (and validate) the accuracy and limitations of each approach.

The project is structured into three main componets: Q1. Traditional Approach: We apply a traditional machine learning algorithm (Decision Tree) to the dataset, explore how well the model performs and how its accuracy can be improved. Q2. Traditional vs. Neural Network Approach: We build and apply a PyTorch neural network to the dataset, comparing its performance to the traditional machine learning model. Q3. Impact of Neural Network Factors: We investigate how different factors (such as pre-training, batch sizes and epochs) affect the performance of the neural network.

Key Objectives:

• Comparison of Algorithms: Compare a traditional machine learning approach (Decision Tree) with a neural network to understand the strengths and limitations of each approach.
• Neural Network Performance: Explore how data handling and hyperparameters affect the performance of a neural network.
• Research Investigation: Address a research question for to neural network optimisation: How does pre-training (and epochs and batches) affect the performance of a neural network?

Tools and Technologies:

• Machine Learning: Scikit-learn for traditional machine learning models.
• Deep Learning: PyTorch for building the neural networks.
• Data Science: Pandas and NumPy for data manipulation; Matplotlib for data visualisation.
• Python Dependencies: A list of dependencies is provided in the dependencies.txt file.

The project includes detailed tutorials for beginners (Q1 & Q2) and more advanced discussions for intermediate users (Q3), with a focus on addressing the reshearch question for NN optimisation.

Dataset Overview

This project provides a detailed exploration of astronomical data from the Sloan Digital Sky Survey (SDSS) Data Release 14 (DR14). The dataset consists of 10,000 observations of taken from the fourth phase of the Sloan Digital Sky Survey (SDSS-IV). Each observation is described by 17 feature columns and 1 target column that identifies the observation as being a star, galaxy, or quasar. The catalog includes object position in the sky (ra and dec), redshift (z) and the observed apparent magnitudes in each filter band, each are crital parameters to training the ML or NN classifiers.

Further details of the data set features can be found at:
SDSS Glossary

To learn more about SDSS:
SDSS Official Website

Source

• Dataset Origin: The data is from the data release DR14 of the SDSS.
• Data Access: There are various ways to gain access to the SDSS data, however more details can be found here:
  SDSS Data Access

This specific dataset was obtained using a sample query from:
SkyServer CasJobs

• Total Rows: 10,000
• Total Columns: 18
• Data Types:
  • 10 columns: float64
  • 7 columns: int64
  • 1 column: object (string)

Column Descriptions

Column	Description	Data Type
objid	Object identifier	float64
ra	Right ascension	float64
dec	Declination	float64
u	Magnitude in the U filter	float64
g	Magnitude in the G filter	float64
r	Magnitude in the R filter	float64
i	Magnitude in the I filter	float64
z	Magnitude in the Z filter	float64
run	Run number of the observation	int64
rerun	Rerun number for calibration	int64
camcol	Camera column number	int64
field	Field number in the run	int64
specobjid	Spectroscopic object identifier	float64
class	Classification of the object (e.g., STAR, GALAXY)	object
redshift	Redshift of the object	float64
plate	Spectroscopic plate number	int64
mjd	Modified Julian Date of the observation	int64
fiberid	Fiber ID for the spectroscopic observation	int64

Sample Data

objid	ra	dec	u	g	r	i	z	class	redshift
1.23765e+18	183.531326	0.089693	19.474	17.042	15.947	15.503	15.225	STAR	-0.00001
1.23765e+18	183.598370	0.135285	18.663	17.214	16.676	16.489	16.392	STAR	-0.00005
1.23765e+18	183.680207	0.126185	19.383	18.192	17.474	17.087	16.801	GALAXY	0.12311
1.23765e+18	183.870529	0.049911	17.765	16.603	16.161	15.982	15.904	STAR	-0.00011
1.23765e+18	183.883288	0.102557	17.550	16.263	16.439	16.555	16.613	STAR	0.00059
1.23765e+18	184.023450	0.073281	20.231	18.634	17.803	17.423	16.992	GALAXY	0.11237
1.23765e+18	184.151283	0.120943	18.951	17.540	16.986	16.698	16.322	STAR	-0.00007

• The dataset includes observed magnitudes for the objects across multiple filter bands (u, g, r, i, z) and metadata for spectroscopic observations with corresponding redshift values.
• Objects are also classified as either STAR, GALAXY or QSO which is the target column the ML and NN models are testest against.

Why Traditional Approaches?

Machine learning has evolved dramatically in recent years, and neural networks have become the go-to solution for many computationlly complex problems. However, traditional machine learning algorithms still have their place, especially when dealing with smaller datasets or problems that don’t require the complexity of deep learning models. In this project, we begin by applying a simple ('easy-to-implement') traditional machine learning method (Desicion Tree) to solve a classification problem. The goal is to establish how well can the traditional method handles the data, and what kind of performance can be expected from a simple model. We also examine the ease of implementation, interpretability, and speed of these models, which remain essential factors in certain scenarios.

Why Neural Networks?

Neural networks today are critical in AI, image recognition and language processing. However, they are often require much more expertise to implement as well more training data and computational power to achieve optimal performance. The motivation for including a neural network in this project is to explore how well these models perform compared to traditional methods when solving the same classification problem. We are particularly interested in understanding how neural networks can overcome limitations that traditional approaches might struggle with, such as non-linear relationships, complex feature interactions, scalability. We also investigate the impact of various factors of neural network architecture (Q3) on model performance.

Dependencies

• numpy == 1.26.4
• pandas == 2.2.2
• matplotlib == 3.10.0
• seaborn == 0.13.2
• scikit-learn == 1.6.0
• pytorch == 2.5.1+cu121
• torchvision == 0.20.1+cu121
• IPython == 7.0.0

Installation and Usage

The repository can be cloned using:

git clone https://github.com/up2013158/SDSS-DR14-Classification-of-Stars-Galaxies-and-Quasars.git

Users who want to clone the repository can also install the dependencies using pip:

!pip install -r dependencies.txt

This command will install all the libraries listed in the dependencies.txt file, ensuring users have all necessary packages need to run the code.

License

This project is licensed under the GNU General Public License v3.0. See the LICENSE file for more details.

References

Acknowledgment for ChatGPT AI:

• Tool Name: ChatGPT
• Provider: OpenAI
• Description: ChatGPT was used to assist in brainstorming, debugging, and generating portions of the codebase for this project.
• Accessed via: https://openai.com/chatgpt

Acknowledgment for Gemini AI:

• Tool Name: Gemini AI
• Provider: Google DeepMind
• Description: Gemini AI was utilised to generate suggestions, optimise algorithms, and enhance the project's overall design.
• Accessed via: Google Colab notebook integration.

Acknowledgment for Claude AI:

• Tool Name: Claude
• Provider: Anthropic
• Description: Claude was employed to generate, debug, and refine code for this project.
• Accessed via: Visual Studio Code extension for Claude AI.

Contact

Samuel Helps

[email protected]

Physics, Astrophysics and Cosmology (Mphys)

University of Portsmouth