prratek / nyctaxi

My work for the New York City Taxi Fare Prediction Kaggle playground competition.

nycTaxi

nycTaxi_rf.ipynb contains my work for the New York City Taxi Fare Prediction Kaggle playground competition. I employ a Random Forest Regressor to predict the fare of a New York City yellow taxi using data that would be available at the start of the ride. Random Forests are a supervised ensemble learning method that bag together several decision trees to correct for their tendency to overfit.

Data: You can find that data for this project from the cometition's Data tab. When running this code, I had train.csv and test.csv in a subdirectory within my working directory called data. If you use a different file structure or change the file names, make sure to update the PATH variable in the notebook and modify the read_csv calls accordingly.

While the goal at the end is to predict our target variable (fare_amount), this notebook pays particular attention to a key stength of the Random Forest – INTERPRETATION. I go through the following general steps:

1. Loading and inspecting the data, cleaning up non-sensical data as needed.
2. Some basic feature engineering using information about the pickup and dropoff locations.
3. Building a quick and dirty Random Forest baseline.
4. Single trees, bagging, and randomness.
5. Datetime feature engineering.

To come:

1. More on hyperparameter tuning.
2. More feature engineering (maybe) - weather, traffic speeds by time, important locations, boroughs, public holidays
3. Interpretation – feature importance, confidence intervals
4. Modular code with separate modules for different functionality that I define in the notebook.

If you have any questions, suggestiong for how this notebook can be improved, or ideas for cool things I can do with Random Forests, feel free to reach out!

Credits: This work is a result of working through the fast.ai Machine Learning 1 course. I have occasionally borrowed code snippets or ideas from the lessons on Random Forests, the code for which can be found at this Github repo.