This repository is the implementation of MCMG.
MCMG is short for 'A Multi-Channel Next POI Recommendation Framework with Multi-Granularity Check-in Signals'. It is a novel framework, equipped with three modules (i.e., global user behavior encoder, local multi-channel encoder, and region-aware weighting strategy), which is designed to achieve the next POI prediction by fusing multi-granularity check-in signals from two orthogonal perspectives, i.e., fine-coarse grained check-ins at either POI/region level or local/global level.
dataset/CAL/AdjacentMatrix_CAL.txt: the adjacent matrix of Calgary;CAL_checkin.csv: check-ins information of Calgary;test_CAL.txt: test set of Calgary;test_group_label_CAL.txt: group lables of trajectories in test set;train_CAL.txt: training set of Calgary;train_group_label_CAL.txt: group lables of trajectories in training set;
NY/NY.rar: all the data files of New York;
PHO/PHO.rar: all the data files of Phoenix;
SIN/SIN.rar: all the data files of Singapore;
appendix.pdf: the best parameter settings for all methods;main.py: main file;model.py: MCMG model file;parameter_setting.py: parameter settings file;utils.py: utils file.
- Environment
- Our proposed MCMG is implemented using pytorch 1.7.1, with Python 3.6.3 from Anaconda 4.3.30. All the experiments are carried out on a machine with Windows 10, Intel CORE i7-8565U CPU, NIVIDA GeForce RTX 2080 and 16G RAM. The following packages are needed (along with their dependencies):
- cuda==11.0
- hyperopt==0.2.4
- numpy==1.19.5
- pandas==1.1.5
- progressbar==2.5
- python==3.6.3
- scipy==1.5.4
- torch==1.7.1
- Our proposed MCMG is implemented using pytorch 1.7.1, with Python 3.6.3 from Anaconda 4.3.30. All the experiments are carried out on a machine with Windows 10, Intel CORE i7-8565U CPU, NIVIDA GeForce RTX 2080 and 16G RAM. The following packages are needed (along with their dependencies):
- Data Preprocessing
- Following state-of-the-arts, for each user, we chronologically divide his check-in records into different trajectories by day, and then take the earlier 80% of his trajectories as training set; the latest 10% of trajectories as the test set; and the rest 10% as the validation set. Besides, we filter out users with less than three trajectories.
$ python main.py (note: use -h to check optional arguments)