E-Commerce Trust Simulation with LLM-Powered Agents

Project Overview

An agent-based modeling (ABM) study simulating an e-commerce marketplace with 7,580 autonomous agents to quantify how coordinated fake-review campaigns manipulate consumer trust and conversion. Built with the MESA framework and local Llama 3.1 8B (via Ollama) for realistic review and purchase behavior, this was the final project for the Simulation and Modeling course at Hasanuddin University.

Research Questions

• RQ1 — How much does a fake-review campaign raise conversion for low-quality products?
• RQ2 — Which consumer persona (Impulsive, Careful, Skeptical) is most vulnerable?

Key Findings

Simulation Architecture

Built on MESA with 5 headphone products (quality tiers from 4.25 to 9.25 out of 10), a random-activation scheduler, and per-iteration data collection. Two low-quality products — BudgetBeats (4.25/10) and ClearSound (5.25/10) — were targeted by fake reviews; three higher-quality products acted as controls.

Agent population (7,580):

• 1,200 genuine reviewers — critical, balanced, and lenient personalities
• 530 fake reviewers — coordinated burst (80 reviews) then maintenance (450 reviews)
• 6,000 shoppers — Impulsive (reads 3 reviews), Careful (10), Skeptical (15)

LLM Integration

Llama 3.1 8B runs locally via Ollama with dynamic context windows (2,048–8,192 tokens) and role-specific temperature (0.3 shoppers, 0.6 genuine reviewers, 0.7 fake reviewers) — at zero API cost. Genuine-reviewer prompts are quality-aware and personality-driven; fake-reviewer prompts always give 5 stars with varied openings to avoid detection patterns; shopper prompts encode persona-specific decision logic.

Experimental Design

Attacks follow a burst + maintenance timeline: baseline (iterations 1–3, genuine reviews only), burst (4–5, 160 fake reviews), then adaptive maintenance (6–20) that scales fake volume to defend the rating (aggressive / moderate / normal). Reviews, transactions, and per-product metrics are exported to CSV each iteration for independent analysis.

RQ1: Fake-Review Impact

Chi-Square tests on the two targeted products:

Both results are extremely significant (p under 0.0001) with large effect sizes (Cramér's V above 0.6). Control products moved only +3–4pp from natural variation, confirming the jumps are caused by the campaign rather than noise.

RQ2: Persona Vulnerability

ANOVA confirms significant differences between personas (F = 540.28, p under 0.0001). Counterintuitively, the Careful persona is most vulnerable (+95pp): reading more reviews backfires when fake reviews dominate the sample. The Skeptical persona — which checks 15 reviews and watches for burst patterns — is 2.3x more resistant (+40pp).

Key Contributions

• First simulation to quantify fake-review conversion lift (+54–72pp) with publication-grade statistical rigor.
• Novel vulnerability finding: deeper-but-naive analysis (Careful) is riskier than skeptical pattern-checking.
• Demonstrates local-LLM agent-based modeling for realistic natural-language behavior at zero API cost.

Practical Implications

• Platforms — flag sudden rating jumps and review bursts; educate the most-vulnerable user segments.
• Consumers — a skeptical mindset and attention to review timing offer 2.3x more protection than trusting star ratings alone.

Performance & Reproducibility

LLM inference averages 1.8s per review (roughly 1,730 reviews in ~52 minutes, ~8 GB RAM); a full 20-iteration run takes ~4–6 hours on consumer hardware, with a fixed random seed (42) and CSV exports for reproducibility.

Repository: fake-review-abm-llm · Course: Simulation and Modeling, Hasanuddin University, 2025