Home

Awesome

HGE: Embedding Temporal Knowledge Graphs in a Product Space of Heterogeneous Geometric Subspaces

This code reproduces results in HGE: Embedding Temporal Knowledge Graphs in a Product Space of Heterogeneous Geometric Subspaces.

Installation

Create a conda environment with pytorch and scikit-learn :

conda create --name tkbc_env python=3.7
source activate tkbc_env
conda install --file requirements.txt -c pytorch

Then install the kbc package to this environment

python setup.py install

Datasets

Datasets are provided in tkbc/src_data folder. Due to file size limit, GDELT will be uploaded in the final version.

Process data by running :

python tkbc/process_icews.py #for icews14, icews05-15 and gdelt
python tkbc/process_wikidata12k.py  # for wikidata12k

This will create the files required to compute the filtered metrics.

Reproducing results

In order to reproduce the results on the datasets in the paper, run the following commands. Detailed hyperparameters could be found in hyperparameter.pdf

python tkbc/learner.py --dataset ICEWS14 --model HGE_TNTComplEx --rank 1200 --emb_reg 3e-3 --time_reg 3e-2 --max_epochs 200  --device cuda:0 --attention 1 --num_form full_rel
python tkbc/learner.py --dataset ICEWS05-15 --model HGE_TNTComplEx --rank 1200 --emb_reg 3e-3 --time_reg 1e-1 --max_epochs 200  --device cuda:0 --attention 1 --num_form full_rel
python tkbc/learner.py --dataset ICEWS14 --model HGE_TComplEx --rank 1200 --emb_reg 3e-3 --time_reg 3e-2 --max_epochs 200  --device cuda:0 --num_form full_rel
python tkbc/learner.py --dataset ICEWS05-15 --model HGE_TComplEx --rank 1200 --emb_reg 3e-3 --time_reg 1e-1 --num_form full_rel --max_epochs 200
python tkbc/learner.py --dataset ICEWS05-15 --model HGE_TLT_KGE --rank 1200 --cycle 1440 --use_reverse --num_form full_rel --max_epochs 200

##Test result

In order to test the results on trained checkpoint, run model_test.py as follows:

python model_test.py --model_dir (checkpoint_path) --device ('cpu' or 'cuda:0')

##Parameters

    '--dataset', type=str, help="Dataset name", ['ICEWS14', 'ICEWS05-15','gdelt','wikidata12k' ]

    '--model', choices=models, models = ['ComplEx', 'TComplEx', 'TNTComplEx', 'HGE_TNTComplEx', 'HGE_TComplEx','HGE_TLT_KGE']
    
    '--max_epochs', default=200, type=int, help="Number of epochs."

    '--valid_freq', default=5, type=int, help="Number of epochs between each valid."

    '--rank', default=100, type=int, help="embedding dimension"
    
    '--batch_size', default=1000, type=int,
    
    '--learning_rate', default=1e-1, type=float,

    '--emb_reg', default=0., type=float, help="Entity/Relation Embedding regularizer weight"
    
    '--time_reg', default=0., type=float, help="Timestamp regularizer weight"

    '--device', default="cpu", help="CPU or cuda:0 device"
    
    '--attention', default=1, type=int, help="use temporal-relational attention or not"

    '--num_form', default="full_rel", help="which space will the model take", select between ['complex','dual', 'split','full_rel'(product space)]

##Ablation Study To test the performance of temporal-relational attention, set --attention = 0.
To test the performance of temporal-geometric attention, set --num_form = 'complex/dual/split' to get results from single geometry.

##Pattern Subset Test To get model's performance on individual temporal pattern set,run:

python instance_test.py

and replaces relevant dataset/checkpoint name in line 246/250

##LCGE Please check LCGE_new folder to see details of LCGE's inference error and our implementation.

License

tkbc is CC-BY-NC licensed, as found in the LICENSE file.

Acknowledge

This project is based on tkbc: https://github.com/facebookresearch/tkbc.