Counterfeit Review Detection (NLP + Risk Scoring)

Tech: Python · scikit‑learn · TF‑IDF · SVM · K‑Means · Isolation Forest · matplotlib / seaborn
Goal: Detect fake reviews and surface high‑risk entities (sellers/categories) for action.

👉 Read the concise one‑pager: docs/one_pager.md

Overview

I built a two‑part pipeline:

1. a fake‑review classifier to score individual reviews; and
2. a Counterfeit Risk Score (CRS) to rank entities using review‑level signals (fake ratio, rating polarisation, burstiness, sentiment‑rating gap, lexical redundancy).

Results (Test Set)

• Accuracy ~90% · Precision ~90.7% · Recall ~89.5% · ROC‑AUC ~96.6%.
• 63 entities evaluated; ~14% Very High and ~43% High risk.
• Examples of flagged entities: Target (Books), Walmart (Beauty), Lowe’s (Sports), HomeDepot (Electronics), Amazon (Automotive).

Visuals

ROC‑AUC by model, final calibrated metrics, confusion matrix, and threshold sensitivity.

Sentiment distribution, histogram of sentiment scores, and sentiment by star rating.

Distributions & relationships for risk indicators; feature correlations.

How CRS is constructed (weights, standardisation, examples) and risk level mapping.

Top high‑risk entities, correlation matrix, category risk spread, and anomaly scatter.

K‑Means cluster characteristics and CRS distribution by cluster.

Isolation Forest results; anomaly vs normal comparison and profiles.

Risk profile lines + detailed metrics table for top entities.

Executive summary with category averages and next‑step recommendations.

How to Run

pip install -r requirements.txt
python code/01_classifier.py
python code/02_counterfeit_risk.py

• Scripts expect data under data/ and will write charts under outputs/.

Repo

.
├── code/                    # Python scripts
├── data/                    # CSV inputs
├── outputs/                 # exported figures
├── docs/                    # one_pager.md
├── requirements.txt
├── LICENSE
└── README.md

Author: Yugal Soni