Test several regression algorithms on Washington bike sharing demand data

This R notebook shows a way to tackle a regression task in R. The data under consideration is the bike sharing demand challenge hosted by kaggle in 2015 (https://www.kaggle.com/c/bike-sharing-demand/). The train and test data is also available at the given URL.

A special focus of this notebook lies on feature engineering and ensemble learning, i.e. finding an optimal linear combination of different models. You will find sections on the following topics:

• Feature engineering guided by some ggplots
• Comparison of very different regression models
• Hyperparameter tuning especially for xgboost and SVMs
• Ensemble learning: Finding a good linear combination of 5 models using nonlinear optimization

Findings

• Hyperparameter tuning improves accuracy significantly.
• As a single model xgboost outperforms (once again) the other models in terms of accuracy and computational time.
• For moderate size of data nonlinear kernel SVMs are still quite competitive.
• The linear combination of models outperforms all single models. The best solution scores at rank 99 (from > 3000 public scores).

For an improved view please use the following nbviewer URL:

https://nbviewer.jupyter.org/github/bschieche/R-bikesharing-demand/blob/master/bikesharing-demand.ipynb