RetraSyn

This is our Python implementation for the paper:

Yujia Hu, Yuntao Du, Zhikun Zhang, Ziquan Fang, Lu Chen, Kai Zheng, Yunjun Gao. Real-Time Trajectory Synthesis with Local Differential Privacy. 2023.

Environment Requirements

• Python >= 3.11.3 (Anaconda3 is recommended)
• numpy == 1.24.3

Dataset

Dataset Statistics

Dataset	Size	# of Points	Average Length	Timestamps
T-Drive	232,640	3,167,316	13.61	886
Oldenburg	260,000	15,597,242	59.98	500
SanJoaquin	1,010,000	55,854,936	55.30	1,000

Dataset Format

For each dataset, there are three corresponding files:

• {datasetname}.xz: contains all trajectory streams in the format of List[List[Tuple[float,float,int]]]

  [Traj1:[(x,y,t),(x,y,t),...],

  Traj2:[(x,y,t),(x,y,t),...],......]

• {datasetname}_transition.xz transforms the original streams into transition states and aligns with the discrete timestamps. The format is List[List[Tuple[float,float,float,float,int]]]:

  [Timestamp1:[(x1,y1,x2,y2,flag)],...],

  Timestamp2:[(x1,y1,x2,y2,flag)],...],......]

  Each 5-tuple represents a transition from (x1,y1) to (x2,y2), flag $\in{0,1,2}$ represents movement transition, entering transition and quitting transition.

• {datasetname}_transition_id.xz is similar to {datasetname}_transition.xz. The difference is the use of a 6-tuple, which add an additional integer to represent user id.

T-Drive

• T-Drive records trajectories of 10,357 taxis in Beijing during one week.
• The dataset is available in ./data/, containing the above mentioned three data files.

Oldenburg

• Oldenburg is a synthetic dataset simulated by Brinkhoff's network-based moving objects generator. It is based on the map of Oldenburg city, Germany.
• Please refer to http://iapg.jade-hs.de/personen/brinkhoff/generator/ to generate the Oldenburg dataset. The parameter used in generation are as follows:
  • obj./begin: 10000 0
  • obj./time: 500 0
  • maximum time: 500
  • classes: 1 0
  • max. speed div.: 50
  • report probability: 1000
• The generated dataset needs to be transformed into the above mentioned data format, containing three data files and located in ./data/

SanJoaquin

• SanJoaquin is based on the map of San Joaquin County, USA.
• Please refer to http://iapg.jade-hs.de/personen/brinkhoff/generator/ to generate the SanJoaquin dataset. The parameter used in generation are as follows:
  • obj./begin: 10000 0
  • obj./time: 1000 0
  • maximum time: 1000
  • classes: 1 0
  • max. speed div.: 50
  • report probability: 1000
• The generated dataset needs to be transformed into the above mentioned data format, containing three data files and located in ./data/

Reproducibility & Run

• Please make sure the data files are located in ./data/ and create directories called ./data/syn_data/{datasetname}/ before running RetraSyn.

• There're two executable files named ./code/RetraSyn_b.py and ./code/RetraSyn_b.py, containing the budget-division strategy and population-division strategy, respectively. Here's an example of running RetraSyn:

  python RetraSyn_b.py --dataset tdrive --grid_num 6 --epsilon 1.0 --w 20 --method retrasyn --multiprocessing
  

  The program will generate a synthesized dataset and save it in ./data/syn_data/{datasetname}/

• To evaluate the synthesized dataset, run ./code/evaluation_b.py or ./code/evaluation_b.py:

  python evaluation_b.py --dataset tdrive --grid_num 6 --epsilon 1.0 --w 20 --method retrasyn --phi 20 --multiprocessing
  

Configurations

The parameters includes:

• --dataset: dataset name, 'tdrive' for default. It should match the name of data files.
• --epsilon: privacy budget
• --grid_num: discretization granularity $K$
• --w: window size
• --phi: size of evaluation time range
• --multiprocessing: whether to use multiprocessing in preprocessing and evaluation
• --method
  • retrasyn: our proposed approach
  • lbd (only in RetraSyn_b.py)
  • lba (only in RetraSyn_b.py)
  • lpd (only in RetraSyn_p.py)
  • lpa (only in RetraSyn_p.py)