Ecosyste.ms: Awesome

An open API service indexing awesome lists of open source software.

Awesome Lists | Featured Topics | Projects

https://github.com/kedwar83/stellar-classification-and-analysis

Star exploration and analysis
https://github.com/kedwar83/stellar-classification-and-analysis

data-science python school-project

Last synced: 3 days ago
JSON representation

Star exploration and analysis

Awesome Lists containing this project

README 

        
# Star Dataset Exploration



## Data



For my second dataset, I will be exploring a dataset about stars. The metrics are:



- **Absolute Temperature (in K)**

- **Relative Luminosity (L/Lo)** (brightness) where Lo = 3.828 x 10^26 Watts (Avg Luminosity of Sun)

- **Relative Radius (R/Ro)** where Ro = 6.9551 x 10^8 m (Avg Radius of Sun)

- **Absolute Magnitude (Mv)** (bigness)

- **Star Color** (white, red, blue, yellow, yellow-orange, etc.)

- **Spectral Class** (O, B, A, F, G, K, M)

- **Star Type** (Red Dwarf, Brown Dwarf, White Dwarf, Main Sequence, SuperGiants, HyperGiants)



## Notes



I believe that this dataset works from the Harvard spectral classification, as well as the Yerkes spectral classification system for luminosity and class respectively. Typically, the Harvard classification system is made up of other variables already present in the dataset, so I am unsure of the efficacy of its inclusion.



## Preprocessing



This dataset has mostly done its own preprocessing. There were no null values. I converted class, color, and type to integers so that they would work nicely with the algorithms. I used `Pipeline` and `StandardScaler` to normalize the data.



## Data Understanding



Some very important notes about this dataset:



- The target class, **star type**, is evenly distributed. This is not actually true in real life, but it's really handy for the data. There are exactly 40 of each kind of star, for zero to five, totaling 240 stars.

- **Spectral class** and **star color** follow something much closer to a logarithmic distribution, with the vast majority of the entries being of the first kind, and the following shrinking by about a factor of two.

- **Temperature**, **luminosity**, **radius**, and **absolute magnitude** are all quite similar in terms of their distribution. They are all high in variability and display moderate clustering, visually speaking.



## Analysis and Results



All models performed extremely well. The performance was so similar that relative comparisons are almost pointless. All models performed in the range of 0.94-1.00, with the Random Forest performing the best and K-Nearest Neighbors (KNN) performing the worst. However, even considering the F1 score and confusion matrix, the models all performed extremely similarly.



Generally, the classification of stars is based pretty simply on the variables in this dataset, so it is not a surprise that all of the models performed so well. A decision tree performing best makes sense, while most of the other models would probably produce overly complex overfit models. Typically, in real-world classification, the process of actually deciding the class of a star is essentially a decision tree.



I do wonder if it would be better to entirely abandon convention and focus on allowing for more complex models based on machine learning techniques. It's hard to know, but that would be neat. If I were to explore that, I would want a less clean dataset with more variables to play with.



## Sources



- [Types of Stars - LCO Global](https://lco.global/spacebook/stars/types-stars/)

- [YouTube Video](https://www.youtube.com/watch?v=Y5VU3Mp6abI)