Feature Augmentation of GNNs for ILPs: Local Uniqueness Suffices

We provide implementations and scripts for three tasks/datasets:

1. ILPs, 2) LPs, and 3) ZINC.

   Each uses a slightly different environment and workflow, detailed below.

Environment Setup

We recommend creating separate environments for the ILP/LP tasks and the ZINC task.

Requirements

To avoid dependency conflicts, create two separate environments: one for ILPs/LPs and one for ZINC.

Scenario	Python	PyTorch	PyG	Other	CUDA
ILPs / LPs	3.8.13	1.10.2	2.0.4	pyscipopt 4.2.0	cudatoolkit 11.3
ZINC	3.11.13	2.7.1	2.6.1	ogb 1.3.6	cudatoolkit 11.8

Note: ogb refers to the Open Graph Benchmark Python package.

Installation Tips

Below are illustrative commands. Choose the correct wheels/indexes for your OS/CUDA combination.

# ILPs / LPs environment
conda create -n ilp python=3.8.13 -y
conda activate ilp
# Install specific versions of torch / pyg as required for your platform
# pip install torch==1.10.2+cu113 ...
# pip install torch-geometric==2.0.4
pip install pyscipopt==4.2.0

  

# ZINC environment
conda create -n zinc python=3.11.13 -y

conda activate zinc

# pip install torch==2.7.1 ...

# pip install torch-geometric==2.6.1

pip install ogb==1.3.6

---

Make sure the CUDA Toolkit version matches your installed PyTorch build.

Data preparation

ILPs: instructions here to prepare the ILPs data.

LP: - generate data by running:

python 1generate_pr.py
python 1generate_wa.py

ZINC: The dataset artifacts are created automatically when you launch the training script.

Training the Model

To train the model , you can use the following bash commands:

ILPs：

epoch=100  
sampleTimes=8  
for dataset in BIP BPP SMSP  
do  
    python train.py --Aug empty --dataset $dataset  --epoch $epoch --sampleTimes $sampleTimes  
    python train.py --Aug uniform --dataset $dataset  --epoch $epoch --sampleTimes $sampleTimes  
    python train.py --Aug pos --dataset $dataset  --epoch $epoch --sampleTimes $sampleTimes  
    python train.py --Aug orbit --dataset $dataset  --epoch $epoch --sampleTimes $sampleTimes  
    python train.py --Aug group --dataset $dataset  --epoch $epoch --sampleTimes $sampleTimes  
    python train.py --Aug color --dataset $dataset  --epoch $epoch --sampleTimes $sampleTimes  
done  
​  
for dataset in BPP  
do  
    python train.py --Aug empty32 --dataset $dataset  --epoch $epoch --sampleTimes $sampleTimes  
    python train.py --Aug colorOrbit --dataset $dataset  --epoch $epoch --sampleTimes $sampleTimes  
    python train.py --Aug colorGroup --dataset $dataset  --epoch $epoch --sampleTimes $sampleTimes  
    python train.py --Aug colorGNN --dataset $dataset  --epoch $epoch --sampleTimes $sampleTimes  
done

LPs:

#WA
python 2trainVanila.py --i color  -p lb
python 2trainVanila.py --i uniform  p lb
python 2trainVanila.py -i vanilla -p lb
#pagerank
python 2trainVanila.py --i color  -p pagerank
python 2trainVanila.py --i uniform  p pagerank
python 2trainVanila.py -i vanilla -p pagerank

ZINC:

python Color_PEARL.py

Evaluation

ILP:Get Top-m% error

statistics regarding Top-m% error can be calculated by running

python read_top_m_error.py

the results will be reported in ./handisTable_valid.xlsx

LP: Get MSE

python 3test.py

ZINC:MAE

Per-seed training logs are saved as:

Color_PEARL_seed{i}_train_log.pkl

where {i} is the seed index you used.