elizabethpeters / mids_girlpower

Modeling Capital Bike Share Demand

Current Standing: 250th

Initial Standing: 1193rd place

Please see FinalProject.ipynb for our final write-up of the code we wrote for the Bike Share competition.

This repo contain the final MIDS Machine Learning 207 course project for Erin Boehmer, Sharon Lin, and Elizabeth Peters. All data was taken from the Kaggle Bike Sharing page and attributed to:

Fanaee-T, Hadi, and Gama, Joao, Event labeling combining ensemble detectors and background knowledge, Progress in Artificial Intelligence (2013): pp. 1-15, Springer Berlin Heidelberg.

mids_girlpower's Issues

Properly represent categorical variables

We had been representing variables like year, month, day, holiday, etc. as float values, when they are not actually continuous. This was mainly to get the baseline to work. So I used the idea of a OneHotEncoder (derived from another code source) to properly encode the variables.

Research how to do Poisson regression in R

Try KNN Method

Bin continuous variables to be more meaningful

Feature engineering of variables like temperature

Play around with binning of temperature/other continuous variables

Look into 3rd party sources for weather

Add a column to data for day_of_week

Use library datetime and function .day(year, month, day) to convert year, month, day variables into days of the week. Then can use the day for prediction.

Create time series graph of registered vs casual rents

I'm thinking a translucent overlapping barchart

Create prediction model for casual rentals

Will need to combine the predictions for this model with the registered prediction to come up with the final count value.