Looking into the past: exploring recurring patterns in time dependencies.
This prject was done in collaboration with Kiril Bikov https://github.com/kiril-bikov for a Geometric Deep Learning submission as part of the MPhil in Advanced Computer Science, University of Cambridge, March 2024.
-
download datasets either manually or through "RC_scripts/TGB_edited/docs/tutorials/Edge_data_numpy.ipynb"
-
adjust pathnames (wd + basepath etc) within rc_main.py and rc_utils.py within RC scripts and run without required arguments:
python rc_main.py
This will both save dataset pickle files and plots if required Plots can also be created separately onces the pickle files are saved through ideas.ipynb and final_plotting.ipynb
-
Within DyGLib_TGB_edited_KB the utils/DataLoader.py has been modified to subselect recurrent interactions.
-
run as an example
python DyGLib_TGB_edited_KB/train_link_prediction.py --dataset_name tgbl-wiki --model_name DyGFormer --max_input_sequence_length 32 --num_neighbors 32 --time_gap 32 --gpu 0 --num_epoch 20 --num_runs 3
-
Move back into RC_scripts and create synthetic dataset using
synthetic_dataset.ipynb. Save the dataset to a 'tgbl-synthetic' folder in at the same level as the other data folders like tgbl-wiki etc. -
Within DyGlib_TFB_rc adjust and run
create_scripts.pyto create the experiments .sh file. -
run
./run_experiments.sh, gradually the results will be placed within a 'logs' and 'saved_results' folders. The DataLoader.py is also edited to save the validation and test predictions for further visualisation within synthetic_dataset.ipynb. These predictions are saved as, for example, 'saved_results/DyGFormer/tgbl-synthetic/val_synthetic_validation_data.csv'.
Follow https://github.com/shenyangHuang/TGB for installation instructions. Base packages will be downloaded using:
pip install py-tgb
To the awesome TGB https://github.com/shenyangHuang/TGB and DyGLib_TGB https://github.com/yule-BUAA/DyGLib_TGB teams for their scripts!!