TGTalker

Official Repository for "Are Large Language Models Good Temporal Graph Learners?"

Installation Instructions

module load python=3.10
python -m venv vllm_env
source vllm_env/bin/activate
pip install vllm

install TGB from source to download datasets and run evaluation

git clone [email protected]:shenyangHuang/TGB.git
pip install -e .   

How to run temporal link prediction with reasoning_main.py

#* base model
CUDA_VISIBLE_DEVICES=0 python -u reasoning_main.py --batch 200 --model qwen1.7b --in_size 5 --bg_size 300 --data tgbl-subreddit --nbr 2 

#* base model + icl
CUDA_VISIBLE_DEVICES=0 python -u reasoning_main.py --batch 200 --model qwen1.7b --in_size 5 --bg_size 300 --data tgbl-subreddit --nbr 2 --icl

#* base model + cot
CUDA_VISIBLE_DEVICES=0 python -u reasoning_main.py --batch 200 --model qwen1.7b --in_size 5 --bg_size 300 --data tgbl-subreddit --nbr 2 --cot --logfile reddit_log.json


#* base model + cot + icl
CUDA_VISIBLE_DEVICES=0 python -u reasoning_main.py --batch 200 --model qwen1.7b --in_size 5 --bg_size 300 --data tgbl-subreddit --nbr 2 --cot --icl --logfile reddit_log.json

How to run temporal link explantion

Generating post-hoc explanations

We assume that:

• We have an answer cache
• We have the prompt cache

Generate Answer Cache

**Run the first 5,000 lines: ** This is the model you want the answer to be from

1. Get answer (e.g.)

   CUDA_VISIBLE_DEVICES=0 python -u reasoning_main.py --batch 200 --model llama3 --in_size 5 --bg_size 300 --data tgbl-subreddit --nbr 2 --icl -—cache_dst

2. Create folder called answer_cache under posthoc_explanations

3. copy {{answer_cache}} to answer_cache

Generate Prompt Cache

**Run the prompt generation ** Keep gpt-4.1-mini-2025-04-14, path currently hard-coded in main.py

1. Get prompts (e.g.)

   python3 ./generate_reasoning_main.py --batch 100 --model gpt-4.1-mini-2025-04-14 --in_size 5 --bg_size 300 --data tgbl-subreddit --nbr 2 --icl --max_no_of_prompts 5000

2. Create folder called output from gpt-batch into posthoc_explanations

Run post-hoc explanation

** This is the model you want the explanation to be generated from + categorization.

python3 ./main.py --data tgbl-subreddit --model llama3

Running openai models

see gpt-batch/README.md for details

Citation:

If this repo is useful for your project, please consider citing our paper:

@article{huang2025large,
  title={Are Large Language Models Good Temporal Graph Learners?},
  author={Huang, Shenyang and Parviz, Ali and Kondrup, Emma and Yang, Zachary and Ding, Zifeng and Bronstein, Michael and Rabbany, Reihaneh and Rabusseau, Guillaume},
  journal={arXiv preprint arXiv:2506.05393},
  year={2025}
}