GSTA

• The source code and data of the paper "GSTA: gated spatial–temporal attention approach for travel time prediction"

Spatial-temporal Attention

Data

• A sample of 81K trips is provided for each of the NYC and Chengdu Taxi datasets in folders (NYC Data, Chengdu Data).
• The data samples are already pre-processed (Data Cleaning, Feature Engineering,... etc) and randomly split into train (Xtrain, Ytrain), validation (Xval, Yval), and test (Xtest, Ytest).
• The implementation of the prediction model for each dataset is given in a separate jupyter notebook (GSTA on NYC.ipynb, and GSTA on Chengdu.ipynb).

Each data sample is a CSV file. The key contains:
* Location Features: 'pickuplongitude', 'pickuplatitude', 'dropofflongitude', 'dropofflatitude', 'centerlatitude', 'centerlongitude', 'dropoffpca0', 'dropoffpca1', 'pickuppca0', 'pickuppca1' * Cluster Features: 'pickupcluster', 'dropoffcluster', 'pickupcountsonclusterid', 'dropoffcountsonclusterid' * Geo-hash Features: 'pickupgeohash', 'dropoffgeohash' * Date/Time Features: 'DayofMonthsin', 'DayofMonthcos', 'Hoursin', 'Hourcos', 'dayofweeksin', 'dayofweekcos', 'Weekendday', 'Workday' * Weather Features: 'tempm', 'dewptm', 'hum', 'rain', 'snow', 'wdird', 'vism', 'fog', 'thunder', 'tornado', 'condsClear', 'condsHaze', 'condsHeavy Rain', 'condsHeavy Snow', 'condsLight Rain', 'condsLight Snow' * Distance: 'distancehaversine', 'distancedummymanhattan' * Direction: 'direction' * Speed: 'avgspeedKMperHour' * Public Holiday: 'PublicHoliday' * Peak period: 'PeakHour' * Travel Time: 'tripduration'

Parameters:

(These parameteres are tuned with whole dataset, you can change them manually)
The default parameters are:
- optimizer=Adam(lr=0.001).
- loss='meanabsoluteerror' - metrics=['mae','mape'] - Dropout=0.2 - epochs = 50 - batchsize = 256 - kernelregularizer=l2(0.001) - Activation('elu') - BatchNormalization(epsilon=1e-06, momentum=0.98) - kernelinitializer="heuniform" - num_heads = 4 , which is the number of heads in Multi-Head attention.

Results

Best Model

The best model for each data during training phase is saved to folder "Models" as hdf5 file.

Dependencies:

Keras 2.4.3, Tensorflow 2.3.0, Bokeh 2.2.1, Numpy 1.19.3, Pandas 1.1.5, Sklearn.

BibTeX Citation

If you use our paper in a scientific publication, we would appreciate using the following citations:

@article{Khaled2021, author = {Khaled, Alkilane and Elsir, Alfateh M Tag and Shen, Yanming}, doi = {10.1007/s00521-021-06560-z}, issn = {1433-3058}, journal = {Neural Computing and Applications}, title = {{GSTA: gated spatial–temporal attention approach for travel time prediction}}, url = {https://doi.org/10.1007/s00521-021-06560-z}, year = {2021} }