Earthquake Rescue Project

In 2010 there was a devastating earthquake in the Caribbean. As part of the rescue effort, a team of data scientists analyzed pixel data of satellite photos for evidence of blue tarps (indicating people in need of assistance).

For a class project I used this problem with the original data to train a variety of classifiers and evaluate performance at different thresholds, utilizing R and R markdown. This project included:

• Extensive EDA

• 10 fold cross validation with the same data split for all models

• Heavily skewed data with around a 3% true positive rate

• ROC and Precision Recall curves to determine ideal thresholds

• Messy, real-world holdout data used for evaluation

• Analytical conclusions based on results of the data

Models considered in the evaluation:

• Logistic regression

• LDA

• QDA

• KNN

• Elastic Net (ridge/lasso)

• Random Forest

• SVM

The clear front runners were logistic regression, LDA, and QDA.

The button below takes you to the full report. To view the full code don’t hesitate to reach out to jordan.hiatt11@gmail.com.