Machine Learning: Exoplanet Exploration

• ML GitHub Link

Background

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.

To help process this data, I created some machine learning models capable of classifying candidate exoplanets from the raw dataset.

In this homework assignment, I performed as:

1. Preprocess the raw data
2. Tuned the models
3. Compared two models

---

Preprocess the Data

• Preprocess the dataset prior to fitting the model.
• Used MinMaxScaler to scale the numerical data.
• Separated the data into training and testing data.

Tune Model Parameters

• Used GridSearch to tune model parameters.
• Tuned and comparedtwo different classifiers.
• Models used were:
  • Logistic Regression (LR)
  • Random Forest Classifier (RFC)

Summary Report

• Models Design:
  • Imported my dependencies as well as loaded the expoplanet_data.csv file.
  • Build both models using all 41 features
  • Instead of deleting columns a priori, I used the base model to evaluate feature importance, and filter the data to include relevant features only.
  • I build a second model by selecting the features model and using the filtered data. *Tuned the model parameters using GridSearchCV.
  • Build the final model using the tuned parameters.
  • Evaluated both models and extracted, as csv and sav files, both Accuracy Report Data Frames.
  • Performed and merged both Accuracy Report Data Frames as a First Glance Comparison Report.

• Models Comparison and Results:
  • The Comparison Report, at first glance, we can see that Random Forest Classifier (RFC) is more accurate than Logistic Regression (LR) by so little margin!
  
  * Eventhough, we can also see that the `Tuned Model` applying the Grid Search CV also refine our accuracy target.
  * Finally, as for Random Forest, we can see that is highly effective applying a feature selection than Logistic Regression model.
  * Conclusions: Given the relatively high accuracy of the RFC model, I believe it to be a reasonable predictor of exoplanet candidacy. However, a model leveraging deep learning techniques might prove superior.

---

Extra Resources

• Exoplanet Data Source

• Scikit-Learn Tutorial Part 1

• Scikit-Learn Tutorial Part 2

• Grid Search

• ML Classification and Regression

• LR vs RFC

---

Hints and Considerations

• Started by cleaning the data, filtering features, and scaling the data.

• Tryed a simple model first, and then tuned the model using GridSearch.

• When hyper-parameter tuning, some models have parameters that depend on each other, and certain combinations will not create a valid model.

• Worked both Models and my Comparison Report in separated Jupyter notebooks, in orde to avoid coding confusion.

---

Submission

• My Jupyter Notebooks for each model are hosted on GitHub.

• Created a file for my best model and push to GitHub

• Included a README.md file that summarizes my assumptions and findings.

• Submitted the link from my GitHub project to Bootcamp Spot.

© 2020 Gabriela Loami Olivares Martinez, BootCamp Tecnologico de Monterrey.

© 2020 Trilogy Education Services, a 2U, Inc. brand. All Rights Reserved.