The following directory contains the Jupyter Notebook for examining the effect of class balancing in cross-validated random forest (CVRF) classifiers. The notebook title ETL.ipynb cleans, reduce and extracts features from the datasets located in the taxi_fare directory. Once the data is ready, the notebook creates, trains, and validates two CVRF's, one with a balanced class dataset and one with the original class imbalance. I use GridSearchCV with k-folds=5 to cross validate each machine learning model. I used parallel computing techinques to run 4 jobs on parallel to reduced run times during the training procedure.

I evaluated the model with two testing sets, the first one was the oringial split done in the notebook, and the second set was obtained from taxi_fare/test.csv which was orignially provided by Kaggle. Both evaluations indicate that class balancing reduced the the number of false negative prediction and increased the number of false positive predictions. Check out the medium post for further discussion: https://medium.com/@patricksandovalromero/a-brief-introduction-cross-validated-random-forest-21423d3378d5

I will create another Medium post where I go into detail on how I am using the pandas library along with the numpy library to reduce my data. However, we can see from the feature corner plot that the 'misc_fees' feature has the clearest patter of distinction for rides that applied and didn't applied surges.

We also see how the 'total_fare' feature shows some clear pattern for discriminating trips with applied surges.

We utilized the K-fold method for cross-validating both models using 5-fold validation meaning we use a 80-20 ratio between the training and validation sets. The tuned hyperparameters for each model are very similar to each other with the exeption of the number_of_estimators, where the balanced CVRF has 75 trees and the imbalanced CVRG has 100.

Using the test set that composed 25% of the original taxi_fare/train.csv dataset we created the following confusion matrix to visualize the propotions of tru positives (TP), true negatives (TN), false positives (FP), and false negatives (FN).

Additionally, we use the mean decrease in impurity as a metric to measure the individual importance and predictive power of features. This allowed us to confirm that misc_fees and total_fare are the most predictive features for surges being applied.

For a more detailed discussion of this project please refer to the medium post reference at the top.