Amazon Co-Purchase Link Prediction

A research implementation of time-aware link prediction on Amazon co-purchase networks with leakage-free temporal evaluation. This project demonstrates rigorous evaluation practices for link prediction on temporal graphs, avoiding common pitfalls like training on future information.

Project Overview

This project builds co-purchase networks from Amazon review data where nodes are products and edges connect products frequently reviewed together. We predict future co-purchases using only historical information through a temporal train/test split at the 80th percentile timestamp (t*), ensuring no information leakage from the future.

The evaluation covers three Amazon product categories (Electronics, All_Beauty, Home_and_Kitchen) and compares classical link prediction baselines under both warm-start (seen nodes) and cold-start (unseen nodes) scenarios.

Authors: Yeseul Han, Erfan YousefMoumji, Xi Lu

Key Contributions

• Leakage-free temporal split: Train on edges up to t* (80th percentile), test on edges after t*
• 2-hop candidate generation: Efficient negative sampling from nodes at distance 2
• Warm vs cold-start evaluation: Separate analysis for seen vs unseen test nodes
• Baseline implementations: Common Neighbors (CN), Jaccard, Adamic-Adar (AA), Preferential Attachment (PA)
• Comprehensive metrics: AUC, Average Precision (AP), Precision@K (K=10,20,50,100,200,500)
• Multi-seed evaluation: 5 random seeds per configuration for statistical robustness
• Runtime profiling: Performance analysis across categories and baselines

Repository Structure

.
├── README.md                      # This file
├── LICENSE                        # MIT License
├── requirements.txt               # Python dependencies
├── CITATION.cff                   # Citation metadata
├── Makefile                       # Build automation (optional)
├── .gitignore                     # Git ignore rules
│
├── src/
│   ├── projectmilestone2.py      # Main pipeline implementation
│   └── run.py                     # Simple entry point 
│
├── data/
│   └── README.md                  # Dataset download instructions
│   # Dataset files go here (not committed)
│
├── results/
│   # CSV files with evaluation metrics
│   ├── 5core_stats.csv
│   ├── eval_linkpred_all.csv
│   ├── eval_linkpred_cold.csv
│   ├── overall_metrics.csv
│   └── coldstart_metrics.csv
│
├── figs/
│   # Key plots and visualizations
│   ├── graph_stats_nodes_edges.png
│   ├── graph_stats_density.png
│   ├── edge_time_cdf_*.png
│   └── ...
│
└── report/
    └── Final_report.pdf           

How to Run

Prerequisites

1. Python 3.7+ with pip installed
2. Download datasets: Follow instructions in data/README.md to download the Amazon 5-core datasets

Installation

# Clone the repository
git clone https://github.com/ErfanYM/amazon-copurchase-link-prediction.git
cd amazon-copurchase-link-prediction

# Install dependencies
pip install -r requirements.txt

Download Data

See data/README.md for detailed instructions. You need:

• Electronics_5.json.gz
• All_Beauty_5.json.gz
• Home_and_Kitchen_5.json.gz

Place them in the data/ directory.

Run the Pipeline

Option 1: Using Make (recommended)

# Install dependencies
make install

# Run the pipeline
make run

# Clean results (if needed)
make clean

Option 2: Using the run script directly

python src/run.py --data-dir data

Option 3: Direct execution (after updating paths in code)

# Edit src/projectmilestone2.py to update data paths (lines 53-55)
# Then run:
python src/projectmilestone2.py

The pipeline will:

1. Load and sample 100K reviews per category
2. Build co-purchase graphs and extract giant connected components (GCC)
3. Compute edge timestamps and determine t* (80th percentile)
4. Split edges into train/test sets (warm and cold-start)
5. Generate 2-hop candidates for link prediction
6. Evaluate 4 baselines × 3 categories × 5 seeds
7. Save results to results/ and plots to figs/

Expected runtime: ~10-30 minutes depending on hardware

Outputs

Results CSVs

• results/5core_stats.csv: Graph statistics (nodes, edges, density, avg degree)
• results/eval_linkpred_all.csv: Per-seed warm evaluation results
• results/eval_linkpred_cold.csv: Per-seed cold-start evaluation results
• results/overall_metrics.csv: Aggregated warm metrics (mean ± std)
• results/coldstart_metrics.csv: Aggregated warm + cold metrics

Figures

• figs/graph_stats_nodes_edges.png: Graph size comparison
• figs/graph_stats_density.png: Density comparison
• figs/edge_time_cdf_*.png: Edge timestamp CDFs with t* cutoff
• Additional plots for AUC/AP/Precision@K comparisons

Dataset

This project uses the Amazon 5-core review datasets (McAuley et al., 2019). The raw data is NOT included in this repository.

• Download: https://nijianmo.github.io/amazon/index.html
• Citation: See data/README.md for proper attribution

Citation

If you use this code or methodology in your research, please cite:

@software{han_yousefmoumji_lu_2026_amazon_linkpred,
  author = {Han, Yeseul and YousefMoumji, Erfan and Lu, Xi},
  title = {Amazon Co-Purchase Link Prediction: Time-Aware, Leakage-Free Evaluation},
  version = {1.0.0},
  year = {2026},
  url = {https://github.com/ErfanYM/amazon-copurchase-link-prediction}
}

See CITATION.cff for machine-readable citation metadata.

License

This project is licensed under the MIT License - see the LICENSE file for details.

Acknowledgments

• Amazon review datasets by Julian McAuley et al.
• NetworkX library for graph analysis
• scikit-learn for evaluation metrics

---

Maintainer: Erfan YousefMoumji
Last Updated: January 2026