ER-N: Robust Recommender Systems Against Coordinated Manipulation

Repository for the ER-N paper. This code demonstrates how ER-N defends against coordinated manipulation attacks where colluding users strategically inflate a target item's ratings.

📋 Prerequisites

• Python 3.7+
• ~500MB disk space (for MovieLens-1M dataset)

🚀 Quick Start (5 minutes)

1. Setup Environment

# Clone repository
git clone https://github.com/nildip/ER-N.git
cd ER-N

# Create virtual environment
python -m venv venv
source venv/bin/activate  # On Windows: venv\Scripts\activate

# Install dependencies
pip install -r requirements.txt

2. Run Quick Test (Synthetic Data)

# Fast test: 1000 users, 200 items, 1000 rounds, 2 seeds
python experiments/run_experiments.py --experiment main --T 1000 --n_seeds 2

Expected output:

Using synthetic: 1000 users, 200 items

Experiment: main_collusion1pct
  Collusion: 1% (10/1000 users)
  ...

ER-N REDUCES MANIPULATION BY: 75.3%

Results saved to: results/main_collusion1pct_synthetic_T1000_seeds2.json

3. Verify Installation

pytest -q

All tests should pass.

---

📊 Reproduce Paper Results

Step 1: Download MovieLens-1M Dataset

bash data/download_datasets.sh

This downloads and extracts MovieLens-1M (~6MB) to data/movielens-1m/.

Step 2: Run Main Experiments

Experiment 1: Main Results (Table 1 in paper)

Tests ER-N vs BSM at different collusion rates.

python experiments/run_experiments.py --experiment main --real-data --K 200 --T 10000 --n_seeds 5

What it does:

• Uses top-200 most popular MovieLens items
• Tests collusion rates: 1%, 5%, 10%, 20%
• 10,000 rounds per seed
• 5 random seeds for statistical validity

Runtime: ~20 minutes

Output files:

• results/main_collusion1pct_movielens_T10000_seeds5.json
• results/main_collusion5pct_movielens_T10000_seeds5.json
• results/main_collusion10pct_movielens_T10000_seeds5.json
• results/main_collusion20pct_movielens_T10000_seeds5.json

---

Experiment 2: Ablation Study - σ (Figure 2 in paper)

Tests how robustness noise parameter affects manipulation resistance.

python experiments/run_experiments.py --experiment ablation-sigma --real-data --K 200 --T 10000 --n_seeds 5

What it does:

• Varies σ: 0.0, 0.1, 0.2, 0.3, 0.5, 0.7, 1.0
• σ=0.0 is equivalent to baseline (no noise)
• Fixed collusion rate: 5%

Runtime: ~30 minutes

Output files:

• results/ablation_sigma0.0_movielens_T10000_seeds5.json
• results/ablation_sigma0.1_movielens_T10000_seeds5.json
• ... (7 files total)

---

Experiment 3: Ablation Study - β (Figure 3 in paper)

Tests exploration parameter sensitivity.

python experiments/run_experiments.py --experiment ablation-beta --real-data --K 200 --T 10000 --n_seeds 5

What it does:

• Varies β: 1.0, 5.0, 10.0, 20.0, 50.0
• Tests exploration vs exploitation trade-off

Runtime: ~25 minutes

Output files:

• results/ablation_beta1.0_movielens_T10000_seeds5.json
• ... (5 files total)

---

Experiment 4: Attack Strength Analysis (Figure 4 in paper)

Tests robustness against varying attack coordination.

python experiments/run_experiments.py --experiment ablation-alpha --real-data --K 200 --T 10000 --n_seeds 5

What it does:

• Varies α (attack strength): 0.0, 0.2, 0.4, 0.6, 0.8, 1.0
• α=0.0: no manipulation
• α=1.0: maximum manipulation

Runtime: ~30 minutes

---

Step 3: Run ALL Experiments

# WARNING: Takes ~2 hours
python experiments/run_experiments.py --experiment all --real-data --K 200 --T 10000 --n_seeds 5

Runs all main + ablation experiments sequentially.

---

Step 4: Generate LaTeX Tables

After running experiments, generate tables for the paper:

python analysis/generate_tables.py

Output:

==================================================================
GENERATING LATEX TABLES
==================================================================

Generating Table 1: Main Results...
✓ Saved to: results/table1_main_results.tex

Generating Table 2: Sigma Ablation...
✓ Saved to: results/table2_ablation_sigma.tex

Generating Table 3: Beta Ablation...
✓ Saved to: results/table3_ablation_beta.tex

Generated tables:

• results/table1_main_results.tex - Main results (Table 1 in paper)
• results/table2_ablation_sigma.tex - Sigma ablation (Table 2 in paper)
• results/table3_ablation_beta.tex - Beta ablation (Table 3 in paper)

Copy these .tex files directly into your LaTeX paper source.

Requirements:

• Must run experiments first (tables need the JSON result files)

---

Step 5: Statistical Significance Tests

Verify that ER-N's improvements are statistically significant:

python analysis/statistical_tests.py

Output:

==================================================================
STATISTICAL TESTS: Main Results
==================================================================

Collusion Rate: 1%
  ER-N boost: 1.12 ± 0.23
  BSM boost:  5.23 ± 0.45
  Mean difference: -4.11
  t-statistic: -15.342
  p-value: 0.0001
  Significant (p<0.05): ✓ YES

Collusion Rate: 5%
  ER-N boost: 3.34 ± 0.67
  BSM boost:  12.45 ± 1.02
  Mean difference: -9.11
  t-statistic: -12.567
  p-value: 0.0002
  Significant (p<0.05): ✓ YES
...

Tests performed:

• Main results: ER-N vs BSM at each collusion rate (paired t-test)
• Sigma ablation: σ=0.0 vs σ=0.3 (validates noise helps)
• Beta ablation: β=1.0 vs β=10.0 (validates exploration helps)

Interpretation:

• p < 0.05 means statistically significant
• Negative mean difference = ER-N has lower manipulation (better)

---

🔧 Advanced Usage

Custom Parameters

# Different collusion rate
python experiments/run_experiments.py --experiment main --collusion 0.15 --T 10000 --n_seeds 5

# Different σ value
python experiments/run_experiments.py --experiment main --sigma 0.4 --T 10000 --n_seeds 5

# Full MovieLens dataset (no top-K filter)
python experiments/run_experiments.py --experiment main --real-data --T 10000 --n_seeds 5

# Synthetic data (fast testing)
python experiments/run_experiments.py --experiment main --T 5000 --n_seeds 3

Parameter Reference

Parameter	Description	Default	Paper Value
--experiment	Which experiment to run	main	varies
--T	Number of rounds	10000	10000
--n_seeds	Random seeds	5	5
--real-data	Use MovieLens	False	True
--K	Top-K items filter	None	200
--beta	Exploration temperature	10.0	10.0
--sigma	ER-N noise parameter	0.3	0.3
--eta0	Learning rate	0.2	0.2
--collusion	Collusion rate	0.05	varies
--alpha	Attack strength	0.8	0.8

---

📁 Project Structure

ER-N/
├── data/
│   ├── download_datasets.sh      # Dataset download script
│   └── preprocess.py             # MovieLens preprocessing
├── src/
│   ├── ern.py                    # ER-N algorithm implementation
│   ├── baselines.py              # Baseline methods (BSM, RobustMF)
│   └── coordinated_manipulation.py  # Attack simulation
├── experiments/
│   └── run_experiments.py        # Main experiment runner
├── analysis/
│   ├── generate_tables.py        # LaTeX table generation
│   └── statistical_tests.py      # Significance testing
├── tests/
│   └── test_*.py                 # Unit tests
├── results/                      # Output JSONs (created automatically)
├── requirements.txt
└── README.md

---

📈 Understanding Results

Each result JSON contains:

{
  "metadata": {
    "collusion_rate": 0.05,
    "n_users": 6040,
    "n_items": 200,
    "T": 10000,
    "beta": 10.0,
    "sigma": 0.3,
    ...
  },
  "results": {
    "ern": [
      {
        "strategic": {
          "target_recs": 523,
          "target_policy_trace": [0.005, 0.007, ...]
        },
        "truthful": {
          "target_recs": 198,
          ...
        },
        "target_boost": 3.25,  // % increase in recommendations
        ...
      }
    ],
    "bsm": [ ... ]
  }
}

Key Metrics:

• target_boost: How much colluders increased target item recommendations (%)
• Lower = better robustness
• ER-N should have ~70-85% lower boost than BSM

---

🧪 Experiment Options

Experiment	Command	Purpose	Runtime
main	--experiment main	Vary collusion rate	~20 min
ablation-sigma	--experiment ablation-sigma	Test σ sensitivity	~30 min
ablation-beta	--experiment ablation-beta	Test β sensitivity	~25 min
ablation-collusion	--experiment ablation-collusion	Extended collusion rates	~30 min
ablation-alpha	--experiment ablation-alpha	Test attack strength	~30 min
all	--experiment all	Run everything	~2 hours

---

🐛 Troubleshooting

Error: "cannot import name 'load_movielens'"

Solution: Run dataset download first:

bash data/download_datasets.sh

Error: "MovieLens not found"

Solution: Check that data/movielens-1m/ratings.dat exists. Re-run download script.

Slow performance

Solution: Use smaller dataset or fewer rounds:

python experiments/run_experiments.py --experiment main --K 100 --T 5000 --n_seeds 3

Out of memory

Solution: Reduce dataset size:

python experiments/run_experiments.py --experiment main --K 50 --T 5000

Missing scipy for statistical tests

Solution: Make sure scipy is installed:

pip install scipy

---

📝 Citation

If you use this code, please cite:

@inproceedings{ern2026,
  title={ER-N: Robust Recommender Systems Against Coordinated Manipulation},
  author={...},
  booktitle={RecSys},
  year={2026}
}

---

📧 Contact

For questions or issues, please open a GitHub issue or contact [your email].

---

🔬 Technical Notes

• Code design: Intentionally simple for clarity and reproducibility
• Scalability: For production use, consider batching and vectorization
• Missing ratings: Filled with global mean (standard imputation)
• Reproducibility: All experiments use fixed random seeds

---

⚡ Quick Reference

# Setup
python -m venv venv && source venv/bin/activate
pip install -r requirements.txt
bash data/download_datasets.sh

# Quick test
python experiments/run_experiments.py --experiment main --T 1000 --n_seeds 2

# Paper results
python experiments/run_experiments.py --experiment all --real-data --K 200 --T 10000 --n_seeds 5

# Generate tables
python analysis/generate_tables.py

# Statistical tests
python analysis/statistical_tests.py

# Single ablation
python experiments/run_experiments.py --experiment ablation-sigma --real-data --K 200 --T 10000 --n_seeds 5