GTA_ACLED_TKG

This is the code to the paper Dynamic Representations of Global Crises: A Temporal Knowledge Graph For Conflicts, Trade and Value Networks We well add the authors after review period has passed.

In the following we describe the steps needed to reproduce our results. It is splot in two parts, 1. Dataset Preprocessing and 2. TKG Forecasting. It is not required to re-run the Dataset Preprocessing Steps. We provide the output of Dataset Preprocessing in /data/crisis2023. These files can be used for TKG Forecasting.

1. Dataset Preprocessing

Sparql Queries:

• Please see README.md in folder queries

Timestep Assignment:

• Run python3 ./data_preprocessing/ts_assignment_gta_star.py and python3 /data_preprocessing/ts_assignment_acled.py to read the .nt files and assign timesteps
• This requires the rdflib package, that can be downloaded here https://github.com/XuguangSong98/rdflib and put into the data_preprocessing folder. Processing data with this package very slow and can take hours to days.
• The output are csv files that can be found in /data/acled and /data/gta respectively

Merge the two datasets:

• Run python3 ./data_preprocessing/merke-tkg-from-gta-acled.py to merge both subsets and create train, valid, test.txt
• What it does:
• Specify timerange of interest. In our case this is 2023-01-01 – 2023-12-31
• Split dataset based on timesteps. Specify train/valid/test split. In our case it is 80/10/10
• Automatically stores the resulting files in /data/crisis2023
• It produces various files:
  • train.txt, valid.txt, test.txt: one line per quadruple, quadruples as subject_id, relation_id, object_id, timestamp (from 0 to num_timesteps), original_dataset_id (0: gta, 1: acled)
  • train_names.txt, valid_names.txt, test_names.txt: one line per quadruple with string description for each node and relation; subject_string, relation_string, object_string, original_dataset_id, timestamp (from 0 to num_timesteps)
  • id_to_node.json and id_to_rel.json: contains dicts with mappings from "node_id" to node_string, and "relation_id" to relation string.
  • node_to_id.json and rel_to_id.json: contains dicts with mappings from node_string to "node_id", and relation string to "relation_id" .
  • stat.txt: two entries, number of nodes, number of distinct relations

2. TKG Forecasting:

All models for TKG Forecasting are in the folder models. Follow the instructions in the respective README.md.

3. Result Evaluation:

The code for evaluating the results for TKG Forecasting are in the folder result_evaluation.py. Follow the instructions in the respective README.md.