floydwch / kddcup2017

A Solution for KDD Cup 2017

Models

The samples are segmented by route with the mean travel time in the time window for the task1, and tollgate/direction with the volume in the time window for the task2. Trained with a blending regression model each task. Besides backward target values(travel time and volume), there are some features related to the sample. See lib/feature.py and task{1,2}/feature.py.

Results

Test1

• Task1: 0.181~0.183 MAPE
• Task2: 0.125~0.131 MAPE

Install Dependencies

pip install -r requirements.txt

Segment Data

python -m task1.segment
python -m task2.segment

Extract Features

python -m task1.extract
python -m task2.extract

Train Models

python -m task1.train
python -m task2.train

Predictions for the Tasks

Since there are test1 and test2 phases, the default is prediction for test1. To predict for the test2 phase, the data should be re-segmented to integrate the test1 data into the training data. The KDD_MODE environment variable should be set to predict to do so (and set to train to switch back to test1).

The resulting files are placed at {task1, task2}/submission/.

Test1

python -m task1.predict
python -m task2.predict

Test2

KDD_MODE=predict python -m task1.segment && python -m task1.extract && python -m task1.train && python -m task1.predict
KDD_MODE=predict python -m task2.segment && python -m task2.extract && python -m task2.train && python -m task2.predict

Customize Models

See task{1,2}/train.py.

Customize Features

See lib/feature.py for common features and task{1,2}/feature.py for specific features, and enable new features in task{1,2}/extract.py.