DeepMove

PyTorch implementation of WWW'18 paper-DeepMove: Predicting Human Mobility with Attentional Recurrent Networks link

Datasets

The sample data to evaluate our model can be found in the data folder, which contains 800+ users and ready for directly used. The raw mobility data similar to ours used in the paper can be found in this public link.

Requirements

Python 2.7
Pytorch 0.20

cPickle is used in the project to store the preprocessed data and parameters. While appearing some warnings, pytorch 0.3.0 can also be used.

Project Structure

/codes
- main.py
- model.py # define models
- sparse_traces.py # foursquare data preprocessing
- train.py # define tools for train the model
/pretrain
- /simple
  - res.m # pretrained model file
  - res.json # detailed evaluation results
  - res.txt # evaluation results
- /simple_long
- /attn_local_long
- /attn_avg_long_user
/data # preprocessed foursquare sample data (pickle file)
/docs # paper and presentation file
/resutls # the default save path when training the model

Usage

Load a pretrained model:

python main.py --model_mode=attn_avg_long_user --pretrain=1

The codes contain four network model (simple, simple_long, attn_avg_long_user, attn_local_long) and a baseline model (Markov). The parameter settings for these model can refer to their res.txt file.

model_in_code	model_in_paper	top-1 accuracy (pre-trained)
markov	markov	0.082
simple	RNN-short	0.096
simple_long	RNN-long	0.118
attn_avg_long_user	Ours attn-1	0.133
attn_local_long	Ours attn-2	0.145

Train a new model:

python main.py --model_mode=attn_avg_long_user --pretrain=0

Other parameters (refer to main.py):

for training:
- learning_rate, lr_step, lr_decay, L2, clip, epoch_max, dropout_p
model definition:
- loc_emb_size, uid_emb_size, tim_emb_size, hidden_size, rnn_type, attn_type
- history_mode: avg, avg, whole

Others

Batch version for this project will come soon.

vonfeng / DeepMove