Winning solution for the Taxi-Trip Time Prediction Challenge on Kaggle (https://www.kaggle.com/c/pkdd-15-taxi-trip-time-prediction-ii).

Documentation about the method is available in doc/Taxi_II_Winning_Solution.pdf. Information on how to generate the solution file can be found below.

I used python 2.7.9 with its popular scientific modules, in particular

• numpy for feature processing,
• pandas for data handling, and
• sklearn for model training.

To download the code, run:

git clone git://github.com/hochthom/kaggle-taxi-ii.git

Download the data files from Kaggle. Place and extract the files into the data directory.

Run the following scripts in any order.

create_training_set_N1.py

Result: train_pp_N1.csv with size: 1658559 x 8, Duration: 679.9 sec *

create_training_set_N2.py

Result: train_pp_N2.csv with size: 1658559 x 10, Duration: 771.8 sec *

create_training_set_N3.py

Result: train_pp_N3.csv with size: 1658559 x 10, Duration: 819.9 sec *

create_training_set.py

Result: train_pp_RND.csv with size: 1928087 x 17, Duration: 911.5 sec *

create_training_set_Experts.py

Result: 319 Training Sets, with size 3 - 109915 x 17, Duration: 10135.2 sec *
train_pp_TST_0.csv - train_pp_TST_319.csv, no file for trip 125.

*Note: The listed running times were achieved on a 12-core server with 24GByte of ram. The running time may be considerably longer if you use a machine with less ram!

mk_submission.py

creates submissions for models N1, N2, N3, RND; Duration: 5175.5 sec *
Result: my_submission_xx.csv with xx in [N1,N2,N3,RND]

mk_submission_Experts.py

creates the expert model predictions, Duration: 2968.3 sec *
Result: predictions_TVT_experts.pkl

*Note: The listed running times were achieved on a 12-core server with 24GByte of ram. The running time may be considerably longer if you use a machine with less ram!

To generate blended prediction files from all the models, run:

blend_submissions.py

Result: final_submission_1.csv and final_submission_2.csv

Submit the files on Kaggle to get it scored.

Submission 1 scored better in the private leaderboard than submission 2 (0.5092 vs. 0.5045), whereas for the public leaderboard it was the other way around (0.5253 vs. 0.5354).