Kaggle NeurIPS 2025 - Open Polymer Prediction: 3rd Solution

This repository contains the complete code and documentation for the winning solution to the Kaggle Open Polymer Prediction competition. The approach is centered around a Graph Attention Network (GATv2) enhanced with task-specific feature engineering and data augmentation.

1. Hardware & Software

• Hardware (Recommended):
  • CPU: Intel(R) Core(TM) Utral 9 185H
  • Memory: 32GB RAM
  • GPU: NVIDIA RTX 4070 ti super
• Software:
  • OS: Linux (Ubuntu 22.04 LTS via WSL2 tested)
  • Environment: Python 3.11 managed with Miniconda
  • Key Libraries: PyTorch 2.6, PyTorch Geometric 2.6, RDKit 2025.3.3

2. Setup and Installation

1. Clone the Repository:

   git clone [your-repo-url]
   cd kaggle_nips_2025_polymer_prediction

2. Create Conda Environment: It is highly recommended to use Conda to manage dependencies.

   # Create a new conda environment named 'polymer_gnn'
   conda create -n polymer_gnn python=3.11 -y
   conda activate polymer_gnn

3. Install Dependencies (IMPORTANT: Multi-step process): Due to build dependencies, the installation must be done in a specific order.

   Step 3.1: Install Core Libraries from requirements.txt First, install PyTorch, RDKit, and other core packages from the requirements.txt file.

   pip install -r requirements.txt

   Step 3.2: Install PyTorch Geometric Libraries Next, install PyG and its dependencies using the official pre-compiled wheels, which avoids build errors. This command installs versions compatible with PyTorch 2.6.0 and CPU. Replace +cpu with your specific CUDA version (e.g., +cu121) if you have a GPU.

   # For CPU-only installation (as tested)
   pip install pyg_lib torch_scatter torch_sparse torch_cluster -f https://data.pyg.org/whl/torch-2.6.0+cpu.html
   pip install torch_geometric

   Note: The command pip install torch_spline_conv from your testing is often included for completeness, but since our model (GATv2) does not use it, installing torch_geometric is sufficient and will pull in necessary dependencies.

4. Download Competition Data: Download the official competition data from the Kaggle competition page. You can use our data in /data/raw/ and processed data in /data/processed/all_augmented_master.pt

3. How to Run the Full Pipeline

The entire solution is executed via three sequential scripts. The commands are also listed in entry_points.md.

IMPORTANT: Ensure you are in the project's root directory (Kaggle_Polymer_GNN_Solution/) when running all the following commands, using the python src/script_name.py format.

Step 1: Prepare Data

This script reads the raw .csv files, performs chemical data augmentation, selects features, and saves the final processed dataset.

python src/prepare_data.py

Step 2: Train Model

This script trains the 5-fold cross-validation models using the processed data.

python src/train.py

Step 3: Generate Predictions

This script loads the 5 trained models and generates the final submission file.

python src/predict.py

• Output: submission.csv will be created in the ./output/ directory.

4. Key Assumptions

• The installation instructions assume a CPU-based environment but can be adapted for GPU by changing the PyG installation command.
• The SETTINGS.json file is the single source of truth for all I/O paths.
• All Python scripts must be executed from the project's root directory.