TSCI (Trajectory-based Social Circle Inference)

This is the implementation of the deep learning based TSCI model: DeepTSCI.

Environment

• python >= 2.7
• Tensorflow 1.0 or ++

Trajectory splitting and POI embedding

• To capture mobility patterns of trajectories, We first split each trajectory into sub-trajectories based on the fixed time interval (see the paper for the detailed explanation).
• Next, we embed POIs into low-dimensional vectors using the common method: wordv2vec.

Dataset

We use four different Location-based Social Network data as follows.

• Gowalla: http://snap.stanford.edu/data/loc-gowalla.html
• Brightkite: http://snap.stanford.edu/data/loc-brightkite.html
• Foursquare(New York,Tokyo): https://sites.google.com/site/yangdingqi/home/foursquare-dataset
• (remark) Please do not use these datasets for commercial purpose. For academic uses, please cite the paper. For Foursquare dataset, we use the same method as in [41] to construct the social networks. Thanks for their help.(see the reference [41] in the paper).

Usage

Using Gowalla data as the example here, similarly to other datasets.
To run LSTM-based DeepTSCI, python GW_LSTM.py/GW_BLSTM.py. The outcome including model and results will be in the folder of out_data.
To run VAE-based DeepTSCI, run the following commands (pre-training and training):

• python GW_VAE.py
• python GW_VAE_S.py

Reference

Hope such an implementation could help you on your projects. Any comments and feedback are appreciated.