README

This is the source code of our papers:

ISWC2020 paper: "Temporal Knowledge Graph Completion based on Time Series Gaussian Embedding"

COLING2020 paper: "TeRo: A Time-aware Knowledge Graph Embedding via Temporal Rotation"

---

Implementation Environment:

• Python 3.7, Pytorch 1.0, CUDA 8.0, Anaconda 4.8.3
• Python 3.7, Pytorch 1.4, CUDA 9.1, Anaconda 4.5.11

Dataset:

• The link of the original dataset YAGO11k can be found from paper: HyTE: Hyperplane-based Temporally aware Knowledge Graph Embedding
• The links of the original datasets ICEWS14 and ICEWS05-15 can be found from paper: Learning Sequence Encoders for Temporal Knowledge Graph Completion
• We uniform the formats of all these datasets.

Usage:

• Install dependencies and put dataset folders here

• model.py contains PyTorch(1.x) based implementation of our proposed models

• To reproduce the reported results of our models, use the following commands:

  python Main.py --model TERO --dataset icews14 --dim 500 --lr 0.1 --gamma 110 --loss logloss --eta 10 --timedisc 0 --cuda True --gran 1
  
  python Main.py --model TERO --dataset icews05-15 --dim 500 --lr 0.1 --gamma 120 --loss logloss --eta 10 --timedisc 0 --cuda True --gran 2
  
  python Main.py --model TERO --dataset yago --dim 500 --lr 0.1 --gamma 50 --loss marginloss --timedisc 2 --cuda True --gran 1 --thre 100
  
  python Main.py --model TERO --dataset wikidata --dim 500 --lr 0.3 --gamma 20 --loss logloss --timedisc 2 --cuda True --gran 1 --thre 300
  
  
  
  python Main.py --model ATISE --dataset icews14 --dim 500 --lr 0.00003 --gamma 120 --loss logloss --timedisc 0 --cuda True --gran 3 --cmin 0.003
  
  python Main.py --model ATISE --dataset icews05-15 --dim 500 --lr 0.00003 --gamma 100 --loss logloss --timedisc 0 --cuda True --gran 30 --cmin 0.003
  
  python Main.py --model ATISE --dataset yago --dim 500 --lr 0.00003 --gamma 1 --loss logloss --timedisc 1 --cuda True --gran 1 --cmin 0.005 --thre 300
  
  python Main.py --model ATISE --dataset wikidata --dim 500 --lr 0.00003 --gamma 1 --loss logloss --timedisc 1 --cuda True --gran 1 --cmin 0.005 --thre 300
  

• Parameters and Some of the important available options include:

    task: [LinkPrediction,TimePrediction]	(default:LinkPrediction)	
    model:  [ATISE,TERO]   (default: ATISE)
    dataset: [icews14,icews05-15,yago,wikidata] (default: icews14)
    max_epoch: (shoud be >500) (default: 5000)
    dim: 	number of dimension (default: 500)
    batch: 	batchsize (default:512)
    lr: 	learning rate (default:0.1)
    gamma: 	margin for translational models (default:1)
    eta:	ratio of negative samples over the positives (default: 10)
    timedisc: the method used for handling facts involving time intervals: 0 means no time intervals; 1 means to discretize time intervals into time points; 2 means to use dual relation embeddings (default: 0)
    cuda:   whether to use cuda devices (default: True)
    loss: use which loss function for optimization: logloss means logistic loss function; marginloss means margin rank loss (default: logloss)
    cmin: minimum threshold of covariance matrices of ATISE (default: 0.005)
    gran: the time unit of icews datasets (default: 1)
    thre: the mini threshold of time classes in yago and wikidata (default: 1)
  

• Results will be printed out and stored in the corresponding dataset folders.

Citation

If you use the codes, please cite the following papers:

• ATiSE:

  @inproceedings{ATiSE,
  title={Temporal knowledge graph completion based on time series gaussian embedding},
  author={Xu, Chenjin and Nayyeri, Mojtaba and Alkhoury, Fouad and Yazdi, Hamed and Lehmann, Jens},
  booktitle={International Semantic Web Conference},
  pages={654--671},
  year={2020},
  organization={Springer}
  }
  

• TeRo:

  @inproceedings{TERO,
    title = "{T}e{R}o: A Time-aware Knowledge Graph Embedding via Temporal Rotation",
    author = "Xu, Chengjin  and
      Nayyeri, Mojtaba  and
      Alkhoury, Fouad  and
      Shariat Yazdi, Hamed  and
      Lehmann, Jens",
    booktitle = "Proceedings of the 28th International Conference on Computational Linguistics",
    month = dec,
    year = "2020",
    address = "Barcelona, Spain (Online)",
    publisher = "International Committee on Computational Linguistics",
    url = "https://aclanthology.org/2020.coling-main.139",
    doi = "10.18653/v1/2020.coling-main.139",
    pages = "1583--1593"
  }
  

License

ATISE is MIT licensed, as found in the LICENSE file.