Estimated Time of Arrival Prediction via Modeling the Spatial-Temporal Interactions between Links and Crosses (GIS Cup)

Competition Page: DiDi-ETA

Feature Generation

Path

python feature.py

• Original train data should be stored at 'GISCUP-submit/data/raw_data/train/'. Original test data should be stored at 'GISCUP-submit/data/raw_data/'
• In 'feature.py'. The variable ws on line 184 represents your workspace, which should be modified to your workspace path
• This script will generate two feature files for each day's data, df and id_ time. They will be located in 'ws/data/feature/df/' and 'ws/data/feature/id_time/', respectively. You don't need to create these directories in advance. If these two directories don't exist, the script will generate them for you
• An additional .pkl file will be created under /id_time directory

Function

• The df feature files contain information of each order and will be used in the W and D module of our model
• There is sequence information in the id_time file. For each sequence, we elaborately insert each cross on a appropriate position into the link sequence
• For the crosses inserted into the link sequence, we assigned new cross_id for them to make their ids and link_id has the same form. cross_dict.npy file records the mapping relationship between old and new IDS
• This script will take a long time to run

Creating the validation set

Path

python val_random_day.py

• In 'val_random_day.py'. The variable ws on line 62 represents your workspace, which should be modified to your workspace path.
• This script will generate two files in 'ws/data/feature/df/' and 'ws/data/feature/id_time/', respectively, which are the new validation set.

Function

• The script will traverse the generated feature files and randomly sample one thirtieth of the data from the feature file of each day as the validation data set.
• The validation set produced will have a similar size and distribution to a single-day training set

Remarks

This script should be run after the feature.py.

Get Results

Path

python train_val_test.py

• In 'train_val_test.py'. The variable ws on line 409 represents your workspace, which should be modified to your workspace path
• After running this file, you can get the result of a single model stored as "model_epoch2_day31.npy" in '~/GISCUP-submit/results

Function

• After running this file, you can get the result of a single model stored as "model_epoch2_day31.npy" in '~/GISCUP-submit/results

• We adopt ensemble averaging to get the final prediction.