🛒 Amazon Product & Customer Behavior Analyzer

A two-axis scoring model that evaluates every product on how much pressure it faces and how much traction it generates — then segments by median-split quadrant, stress-tests, and reports actionable findings.

📊 Dashboard Preview

(Run the notebook to generate all charts — examples shown from a live run on 1,465 Amazon products across 2 categories)

Dashboard	Quadrant Map
6-panel overview with segment distribution, PVS histogram, sensitivity heatmap	Pressure × Traction quadrant scatter with median thresholds

Discount–PVS Relationship	Word Clouds
OLS regression of discount % vs Perceived Value Score by segment	Per-segment word clouds (Champions / Rising Stars / Core / Vulnerable)

---

🔍 Scoring Framework

Axis	Name	Measures
Axis 1	Pressure Index	Discount dependency · rating weakness · low social proof
Axis 2	Traction Index	Engagement depth · quality signals · sentiment
Combined	Product Vitality Score (PVS)	Weighted blend — the north-star metric

All components are percentile ranks — the model grades products relative to each other so output is meaningful regardless of whether the dataset skews premium or budget overall. This guarantees mean ≈ 0.500 by mathematical construction.

Pressure Index (PI)

PI = 0.40 × discount_rank + 0.40 × rating_risk_rank + 0.20 × low_engagement_rank

Traction Index (TI)

TI = 0.35 × engagement_rank + 0.35 × category_adj_rating_rank + 0.30 × positive_sentiment

Product Vitality Score (PVS)

PVS = 0.50 × Traction + 0.30 × category_adj_rating + 0.20 × sentiment

---

🗺️ Segmentation

Segments are assigned by rule-based median split — directly from each product's Pressure × Traction position relative to the dataset medians. This is an explicit, interpretable decision boundary — not a claim about natural groupings in the data.

Segment	Quadrant	Strategy
🏆 Champions	High traction · Low pressure	Loyalty programme · cross-sell
🌱 Rising Stars	High traction · High pressure	Reduce discount dependency · generate reviews
📊 Core	Low traction · Low pressure	Stable baseline — monitor
⚠️ Vulnerable	Low traction · High pressure	Fix quality issues first — discount cuts won't fix traction

Why not a data-driven clustering method?
K-Means and GMM were both evaluated. Neither found four natural clusters in this data (best silhouette ≈ 0.12–0.26). Amazon product data on Pressure × Traction forms a continuum, not discrete groups. A median-split is scientifically more honest: it is an explicit editorial choice rather than a spurious claim of natural separation.

K-Means (k=4) is still run in the background only to derive a per-product confidence score — how cleanly a product sits near its quadrant centre vs. the other three. Segment labels come exclusively from the median-split logic.

---

📈 Sample Results (1,465 products · 2 categories)

SCORES  (percentile-rank — mean ≈ 0.500 by design)
  Pressure:    0.500  Std=0.193
  Traction:    0.504  Std=0.162
  PVS:         0.505  Std=0.164

SEGMENTS  (rule-based median split — ≈ 25 % per quadrant by design)
  🏆 Champions:    573 (39.1%)   hi traction · lo pressure
  🌱 Rising Stars: 160 (10.9%)   hi traction · hi pressure
  📊 Core:         159 (10.9%)   lo traction · lo pressure
  ⚠️  Vulnerable:  573 (39.1%)   lo traction · hi pressure

SEGMENT QUALITY
  Silhouette:      0.1230
  ANOVA p-value:   0.00e+00  ✅ Significant
  Mean confidence: 0.743

SUSPICIOUS REVIEWS
  Total flagged:   102 (7.0%)

CATEGORY INSIGHTS
  Categories:      Computers & Accessories · Electronics
  Highest pressure: Electronics

SENSITIVITY
  Strongest lever:  Rating quality push  (+10.0% PVS)
  Riskiest action:  Discount +10%  →  PVS -2.6%

---

🚀 Quick Start

Option A — Google Colab (recommended, zero setup)

1. Open Amazon_Customer_Behavior_Analyzer.ipynb in Google Colab
2. Run All (Runtime → Run all)
3. The notebook auto-downloads the dataset via kagglehub
4. Run Section 18 to download all output files

Option B — Kaggle

1. Create a new Kaggle Notebook
2. Add dataset: Amazon Sales Dataset via + Add Data
3. Upload Amazon_Customer_Behavior_Analyzer.ipynb
4. Run All — output files appear in the right-side panel

Option C — Local

pip install vaderSentiment kagglehub wordcloud plotly scikit-learn textblob pandas numpy matplotlib seaborn scipy statsmodels
jupyter notebook Amazon_Customer_Behavior_Analyzer.ipynb

---

📦 Output Files

File	Description
results.csv	Full scored dataset with all metrics
quadrant_map.png	Pressure × Traction segment scatter
score_distributions.png	Score calibration check
segment_quality.png	Silhouette plot per segment
price_analysis.png	Discount % vs PVS OLS regression
pca.png	2D PCA projection of all features by segment
wordclouds.png	Most frequent product terms
rfm.png	Engagement quintile tier distribution
05b_sensitivity_heatmap.png	Δ Mean PVS per segment under stress-test scenarios
09_category_heatmap.png	Category × metric heatmap
10_rising_stars.png	Rising Stars by category
interactive.html	Plotly interactive dashboard (open in browser)
interactive_dark.html	Dark-theme version of the dashboard
suspicious_products.csv	Anomaly-flagged products

---

🔬 Methodology

Why percentile ranks?

Absolute scores are dataset-dependent — a "70% discount" means something different in electronics vs. office products. Percentile ranks normalize across the entire catalogue so every score is relative to peers. The mathematical consequence is mean ≈ 0.500 for all axes — a useful sanity check.

Segmentation methodology

Segmentation uses a median split on Pressure and Traction. This is a deliberate, rule-based approach:

• Transparent: the boundary is the dataset median — easy to explain and reproduce
• Stable: segment sizes are guaranteed to be roughly equal (≈25% each quadrant)
• Honest: the data does not contain four natural clusters (silhouette ≈ 0.12); claiming otherwise would be misleading

The PVS score (Perceived Value Score) is the continuous north-star metric. Segmentation provides an interpretable summary on top of PVS — not a replacement for it.

Sentiment pipeline

VADER (50%) + TextBlob (50%) → raw ensemble → Z-score calibration → clip(-3,3)/3.0

Z-score calibration ensures mean ≈ 0 regardless of how uniformly positive Amazon reviews are. Falls back to a rating-based proxy (rating − 3) / 2 when fewer than 100 products have usable review text.

Anomaly-flagged products

Isolation Forest (contamination=7%) detects statistical anomalies in the review-count × discount × rating × engagement feature space. Flagged products are saved to suspicious_products.csv. PVS scores for flagged products should not be used as-is without manual audit.

Engagement Quintile Tiers (RFM-proxy)

The notebook segments products into five engagement tiers using quintile scoring on:

• R — inverse discount dependency (low discount = organically demanded)
• F — log-normalised review count (interaction volume)
• M — actual_price × rating_norm (perceived revenue contribution)

These are product-level proxies, not true customer-level RFM — the dataset does not contain transaction timestamps.

Discount–PVS relationship

OLS regression of discount_pct on PVS quantifies how much each percentage point of discount correlates with perceived value. The β coefficient and Pearson r are annotated directly on the scatter plot.

---

📁 Repository Structure

Amazon_Customer_Behavior_Analyzer/
├── Amazon_Customer_Behavior_Analyzer.ipynb   # Main analysis notebook
├── README.md                                  # This file
├── LICENSE                                    # Apache 2.0
└── .gitignore

---

🗂️ Dataset

Amazon Sales Dataset by Karavel Raja (Kaggle)

The notebook auto-detects column roles by name pattern — it will work with any CSV that has columns containing the words rating, discount, category, price. No manual configuration required.

---

📋 Requirements

pandas >= 1.3
numpy >= 1.21
matplotlib >= 3.4
seaborn >= 0.11
plotly >= 5.0
scikit-learn >= 0.24
scipy >= 1.7
wordcloud >= 1.8
vaderSentiment >= 3.3
textblob >= 0.17
statsmodels >= 0.13
kagglehub >= 0.1

---

📄 License

Apache License 2.0 — see LICENSE for details.