Fake Review ABM Simulation

Agent-Based Modeling (ABM) simulation untuk meneliti dampak fake reviews terhadap keputusan pembelian di e-commerce menggunakan Large Language Models (LLM).

Research Questions

• RQ1: How much does conversion rate increase for low-quality products targeted by fake review campaigns?
• RQ2: Which consumer persona is most vulnerable to fake reviews? (Impulsive vs Careful vs Skeptical)

Key Features

• MESA Framework: Python-based ABM with RandomActivation scheduler
• LLM-Powered Agents: Natural language reviews and decisions via Llama 3.1 8B
• Publication-Grade Analysis: Chi-Square, ANOVA, Cramér's V effect sizes
• Realistic Attack Scenario: Burst + maintenance fake review campaigns

Technology Stack

Component	Technology
ABM Framework	MESA (Python)
LLM Engine	Ollama + Llama 3.1 8B
Data Storage	In-memory + CSV exports
Analysis	Pandas, NumPy, SciPy
Visualization	Matplotlib

Project Structure

fake_review_sim/
├── src/
│   ├── config.py          # Simulation parameters
│   ├── model.py           # MESA model
│   ├── agents.py          # Reviewer & Shopper agents
│   ├── prompts.py         # LLM prompt templates
│   └── llm_client.py      # Ollama wrapper
├── analysis/
│   ├── csv_analysis.py          # CSV-based analysis
│   └── publication_figures.py   # Publication figures generator
├── data/
│   ├── results/           # Simulation outputs (CSV)
│   └── publication/       # Figures & tables
├── main.py                # Entry point
├── requirements.txt
└── README.md

Installation

1. Install Ollama

Download from: https://ollama.com/download

ollama --version

2. Download LLM Model

ollama pull llama3.1:8b

3. Setup Python Environment

python -m venv venv
.\venv\Scripts\activate
pip install -r requirements.txt

Usage

1. Run Simulation

python main.py

Runtime: 4-6 hours (overnight recommended)

2. Generate Publication Figures

python analysis\publication_figures.py

Outputs (in data/publication/):

• Figure_1_RQ1.pdf - Fake review impact
• Figure_2_RQ2.pdf - Persona vulnerability
• Figure_3_Temporal.pdf - Temporal dynamics
• Table_1_Products.tex - Product characteristics
• Table_2_RQ1.tex - RQ1 statistical results
• Table_3_RQ2.tex - RQ2 statistical results

Configuration

From src/config.py:

TOTAL_ITERATIONS = 20
GENUINE_REVIEWS_PER_PRODUCT = 12  # 4 Critical + 4 Balanced + 4 Lenient
BURST_ITERATIONS = [4, 5]
BURST_VOLUME = {4: 40, 5: 40}
MAINTENANCE_VOLUME = 15
SHOPPERS_PER_PRODUCT_PER_PERSONA = 20
FAKE_CAMPAIGN_TARGETS = [3, 5]  # BudgetBeats, ClearSound Basic

REVIEWS_READ = {
    "Impulsive": 3,
    "Careful": 10,
    "Skeptical": 15
}

Products

ID	Product	Quality	Price (IDR)	Targeted
1	SoundMax Pro	High	450,000	No
2	AudioBlast Wireless	Medium-High	350,000	No
3	BudgetBeats	Low	150,000	Yes
4	TechWave Elite	Premium	650,000	No
5	ClearSound Basic	Low-Medium	250,000	Yes

Agent Types

Reviewers

Type	Personality	Rating Tendency
Genuine	Critical	-1 star bias
Genuine	Balanced	Objective
Genuine	Lenient	+1 star bias
Fake	-	Always 5 stars

Shoppers

Persona	Reviews Read	Vulnerability
Impulsive	3	HIGH
Careful	10	MEDIUM
Skeptical	15	LOW

Results

RQ1: Fake Review Impact

Product	Baseline	Attack	Increase	χ²	p-value
BudgetBeats	0%	54%	+54 pp	121.30	<0.0001***
ClearSound	0%	72%	+72 pp	177.35	<0.0001***

RQ2: Persona Vulnerability

Rank	Persona	Impact (pp)
1	Impulsive	+75.0
2	Careful	+72.5
3	Skeptical	+41.2

Temporal Dynamics

Attack Timeline

Iteration:  1   2   3   4   5   6   7   8  ...  20
            |-------|   |---|   |---------------|
            BASELINE    BURST     MAINTENANCE

Baseline (1-3):   Only genuine reviews
Burst (4-5):      40 fake reviews/iteration
Maintenance (6+): 15 fake reviews/iteration

Statistical Methods

• Chi-Square Test: Compare conversion rates (baseline vs attack)
• Cramér's V: Effect size for Chi-Square
• ANOVA: Compare conversion rates across personas

Troubleshooting

Ollama Not Responding

ollama list

Slow Simulation

• Check GPU usage: nvidia-smi
• Normal speed: ~2-3 seconds per LLM call

ModuleNotFoundError

.\venv\Scripts\activate
pip install -r requirements.txt

License

Educational/Research use only.

Citation

@software{fake_review_abm_2025,
  author = {Ikrar},
  title = {Fake Review ABM Simulation: LLM-Powered Agent-Based Modeling},
  year = {2025},
  publisher = {GitHub}
}

Author

Ikrar Gempur Tirani - D121231015 Muhammad Irgi Abayl Marzuku - D121231102 Ahnaf Fauzan Zaki - D121231033