Source code for our paper entitled "Deep Transfer Learning for Intelligent Wireless Traffic Prediction Based on Cross-Domain Big Data"

The datasets are available in folders of Github_Version/data/. Feel free to download it and test our algorithm.

see our paper for more details.

How to download our shared data (new)

For someone that can not download our data from this repo, please find it at google drive: https://drive.google.com/file/d/1ITSRwdGOra1wZxeZyC4d48_AYvmAcJeE/view?usp=sharing

Note that if you use git clone command to clone this project, the dataset maybe not properly downloaded. If this is the case for you, please go to the directory STCNet/Github_Version/data/data_git_version.7z and click view raw, this operation will manually download the dataset.

Or maybe you need to install Git LFS to sucessfully clone this project.

Script that generates the H5 file is included

As there are so many people that wrote to me asking how to generate the H5 file. So I made this avaliable for other people convience. Please find the code at the Github_Version/data/.

Requirements

System: Ubuntu 16.04 LTS with GeForce GTX TITAN X, 64 bit OS

Python: 3.6.3 Anaconda 64-bit

Pytorch version: 0.4.0

CUDA Driver version: 384.130

About the data

The original data is with precision of 15 decimal digits and the size is about 500 MB. To reduce the data size so that it can be downloaded by others, the data is rounded to 3 decimals. That is, the data is manipulated through

data = np.around(data, 3)

Besides, the original data has five channels, i.e., (sms_in, sms_out, call_in, call_out, internet). We grouped them as (sms, call, internet), that is, three channels.

The wireless traffic data is named as "data_git_version.7z", please unzip it to your local computer.

The crawled cross-domain data is named as "crawled_feature.csv", there are 4 columns, i.e., (social, BSs, POI1, POI2).

There is also a file named as "cluster_label_20.csv", this is the clustered results.

How to execute

Download the code to your own computer, open the terminal, simply type:

python demo_three_cluster.py

You can also specify parameters to control the training, such as set "epoch_size" to 300

python demo_three_cluster.py -epoch_size 300

There are others parameters can be set, please see the code for details.

Note

As the last week of this dataset contains the New Year's Eve, so for that specific data point, we do not predict it as it is a very challenging task to predict such "outliers". We use a simple linear prediction, that is, $y_t = (y_{t-1}+y_{t-2}+y_{t-3})/3*n$.