ReStruct

Large Language Model-driven Meta-structure Discovery in Heterogeneous Information Network

Environment

python==3.8
pytorch==1.13.1 with CUDA support
numpy==1.23.5
networkx==2.8.8
scikit-learn==1.2.0
pandas==1.5.2

File description

preprocess_recommendation.py: Script that preprocesses data for recommendation.

gen_neg_recommendation.py: Script that preprocesses data for recommendation (generates negative samples).

preprocess_node_classification.py: Script that preprocesses data for node classification.

data_recommendation/: Preprocessed data for recommendation.

data_node_classification/: Preprocessed data for node classification.

llm_component_no_prob.py: LLM object.

main_recommendation.py: Script that searches for meta-structures for recommendation.

main_node_classification.py: Script that searches for meta-structures for node classification.

model_recommendation.py: Model for evaluation in recommendation.

model_node_classification.py: Model for evaluation in node classification.

train_recommendation.py: Script that evaluates discovered meta-structures for recommendation.

train_node_classification.py: Script that evaluates discovered meta-structures for node classification.

llm_explanation.py: Script that generates differential explanations for meta-structures.

utils.py: Script that contains useful functions.

Preparing data

**We already provide the preprocessed data under data_recommendation/ and data_node_classification/. **

Search for meta-structures (example)

python main_recommendation.py --dataset Yelp --num_generations 30
python main_node_classification.py --dataset IMDB --num_generations 30

Evaluation (example)

python train_recommendation.py --dataset Yelp --lr 0.01 --wd 0.002 --dropout 0.5  
python train_node_classification.py --dataset IMDB --lr 0.01 --wd 0.0002 --dropout 0.75 --no_norm