MST-OATD

Code for "Multi-Scale Detection of Anomalous Spatio-Temporal Trajectories in Evolving Trajectory Datasets"

Requirements

pip install -r requirements.txt

Preprocessing

• Step1: Download the Porto dataset (train.csv.zip) from Porto, and the Chengdu dataset (Chengdu.zip) from Chengdu.
• Step2: Put the Porto data file in ../datasets/porto/, and unzip it as porto.csv. Put the unzipped Chengdu data in ../datasets/chengdu/.
• Step3: Run preprocessing by

  mkdir -p data/<dataset_name>
  cd preprocess
  python preprocess_<dataset_name>.py
  cd ..
  mkdir logs models probs

  : porto or cd

Generating ground truth

python generate_outliers.py --distance 2 --fraction 0.2 --obeserved_ratio 1.0 --dataset <dataset_name>

distance is used to control the moving distance of outliers, fraction is the fraction of continuous outlier, obeserved_ratio is the ratio of the obeserved part of a trajectory.

Training and testing

python train.py --task train --dataset <dataset_name>
python train.py --task test --distance 2 --fraction 0.2 --obeserved_ratio 1.0 --dataset <dataset_name>

Training on evolving datasets

python train_labels.py --dataset <dataset_name>
python train_update.py --update_mode pretrain --dataset <dataset_name> --train_num <train_num>

update_mode contains three modes: pretrain, temporal, rank, is the number of trajectories used for evolving training.

Citation

Please kindly cite our work if you find our paper or codes helpful.

@inproceedings{wang2024multi,
  title={Multi-Scale Detection of Anomalous Spatio-Temporal Trajectories in Evolving Trajectory Datasets},
  author={Wang, Chenhao and Chen, Lisi and Shang, Shuo and Jensen, Christian S and Kalnis, Panos},
  booktitle={Proceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining},
  pages={2980--2990},
  year={2024}
}