oosmetricsTop 10 in Training by acceleration - 2026-04-27

Kaggle NeurIPS – Open Polymer Prediction

Reproducible and Improved Silver-Medal Solution

This repository is an engineering-refined and fully reproducible implementation based on
Kaggle NeurIPS 2025 – Open Polymer Prediction (Silver Medal) solution.

The pipeline combines:

• SMILES canonicalization and data cleaning (RDKit)
• Fingerprint + descriptor feature engineering
• Group-aware cross-validation (to avoid data leakage)
• Traditional ML models (CatBoost, XGBoost)
• Optional GNN model (PyTorch Geometric)
• Out-of-Fold (OOF) stacking and ensemble learning
• Optuna-based weight optimization (optional)

The goal is to provide:
✅ higher stability
✅ better generalization
✅ strict reproducibility
✅ clean project structure

📂 Project Structure

Kaggle-NeurIPS-Open-Polymer/
├── README.md
├── requirements.txt
├── config.py
├── dataset.csv
├── src/
│ ├── train.py
│ ├── features.py
│ ├── models.py
│ ├── utils/
│ │ ├── smiles_utils.py
│ │ └── cv_utils.py
│ └── stacking/
│ └── optuna_stack.py
├── outputs/
│ ├── models/
│ ├── feats/
│ └── oof/
└── tests/
└── test_smiles.py

⚙️ Environment Setup (Recommended)

We strongly recommend using Conda to install RDKit.

1. Create environment

conda create -n polymer_env python=3.10 -y
conda activate polymer_env

2. Install RDKit

conda install -c conda-forge rdkit -y

3. Install remaining dependencies

pip install -r requirements.txt

🔬 Pipeline Overview

1. SMILES Canonicalization

•	Uses RDKit to convert SMILES into canonical isomeric form
•	Invalid SMILES are logged to outputs/oof/failed_smiles.csv

2. Feature Engineering

•	Morgan fingerprints (2048 bits)
•	Truncated SVD for dimensionality reduction
•	RDKit descriptors (MolWt, TPSA, LogP)

3. Cross Validation

•	GroupKFold based on canonical SMILES
•	Prevents molecule-level leakage

4. Models

•	CatBoostRegressor
•	XGBoostRegressor
•	Optional GNN (PyTorch Geometric)

5. Ensemble

•	Ridge regression as meta-learner
•	Optional Optuna-optimized weighted ensemble

6. Reproducibility

•	Fixed random seed
•	Cached features
•	Saved models and OOF predictions

🚀 Quick Start (Demo Mode)

Run a small synthetic dataset to verify installation:

python src/train.py --demo

This will:
• Generate a fake dataset
• Train models
• Output results to outputs/

🧩 Extending the Project

You can easily add:
• New molecular descriptors
• New GNN architectures (GAT, GraphSAGE)
• New meta learners (LightGBM, ElasticNet)
• Bayesian uncertainty estimation
• Multi-task prediction heads