Multi-Attention-Spatiotemporal-Network-for-mobile-traffic-prediction

Implementing Multi-level attention spatio-temporal network for mobile traffic prediction using Keras.

Download the data

Mobile traffic data is released by Telecom Italia and the data can be acquired here.

You can download the first 7 days of November as demo for testing the code.

Cleaning data and making data set

data preprocesing and generating h5 File

• run : src> python datapreprocessing.py ../data/raw ../data/processed

making dataset

• to calculate series distance, run: src> python calSeriesDis.py

  notice that series distance is calculated based on weekly average traffic series, so if using only 7 days as demo for testing, the whole demo data will be used to calculate series distance

• to generate weight matrix, run: src>python genweight.py

• to train and test data set for the model, run : src >python makedataset.py

  test_len is specified in this py file

MASTNN Model

Framework

Train and evaluate model

• src> python train.py

  model type is specified by the parameter modelbase

• baseline model STN(spatio-temporal network that incoporating 3Dconv and convlstm for forecasting)$[3]$

  • make dataset: src > python stn_makedataset.py
  • train and evaluate: src > python stn_model_train.py

• baseline mode ARIMA:

  • run: src > python arima_train_evaluate.py

Reference

• 1.Barlacchi G , De Nadai M , Larcher R , et al. A multi-source dataset of urban life in the city of Milan and the Province of Trentino[J]. Scientific Data, 2015, 2:150055.
• 2.Yuxuan Liang, Songyu Ke, Junbo Zhang, Xiuwen Yi, Yu Zheng, "GeoMAN: Multi-level Attention Networks for Geo-sensory Time Series Prediction", IJCAI, 2018.
• 3.Zhang C , Patras P . Long-Term Mobile Traffic Forecasting Using Deep Spatio-Temporal Neural Networks[J]. 2017.