Created supervised machine learning models capable of classifying candidate exoplanets from the raw dataset. Conducted feature selection and data scaling. Two classification models, Support Vector and Neural Network, were trained, hypertuned, and tested.
Over a period of nine years in deep space, the NASA Kepler space telescope has been out on a planet-hunting mission to discover hidden planets outside of our solar system, otherwise known as "exo-planets". The dataset included just under 7000 records classified as either "confirmed", "false positive", or "candidate".
Data preprocessing entailed 1) data cleaning, 2) feature selection, and 3) data scaling.
I dropped any rows with missing values and removed those classified as "candidates" to ensure the model learned on the records that had been correctly classified. The final dataset had 5,304 records.
Feature selection allows us to use only the features that have a sizable impact on the classification. Reducing the features in the model allows for 1) quicker model training, 2) a less complex model, and 3) a reduction in overfitting the model.
I created a correlation matrix heatmap to get a visual of how the features correlated to the outcome variable and each ohter.
I used the sklearn library to give each feature a score to show its importance in classification. Here we can see that KOI_depth variable has the highest score and is the most imporant feature. Based on the scrores I decided to use the top seven features in my model.
The two classification models used were Support Vector and Neural Networks. The models were trained, tested, and hyptertuned to compare the performance. Support vector had an Accuracy score of .8 and an F1 score of .81. The neural netork model had similar scores: Accuracy = .81, F1 score = .81. Below is the confusion matrix for the support vector model:
I wanted to understand how the support vector model would perform with all 40 features included even though there is the potential for an overfitting of the model. The scores went up significantly! Accuray = .99, and F1 score = .99. Below is the support vector confustion matrix with all 40 features:
The Kaggle explanet dataset used can be found here: dataset
Michael Occhicone, [email protected]