Executive Summary

The Challenge: Previous clustering attempts failed because physical dimensions (weight, volume) and logistics failures drowned out true product quality. A 1-star review due to a 30-day shipping delay was incorrectly penalizing the product itself.

The Solution: We engineered a pipeline that strictly isolates pure product dissatisfaction from logistics failures. We only evaluate reviews for orders that were delivered on-time or early.

The Result: A clean, actionable segmentation separating our flawless premium products from our high-defect liabilities.

Data Engineering: Finding the “True” Metrics

To find the actual customer experience, we engineered two core metrics:

1. True Defect Rate: % of 1-star and 2-star reviews (On-time deliveries only).
2. Average Quality Score: Mean review score 1-5 (On-time deliveries only).

Note: All physical product dimensions and shipping dates were explicitly dropped from the model after calculating logistics delay to prevent distance matrix corruption.

The Baseline: Algorithmic Clustering

We first run Hierarchical Clustering (Ward’s Method) using our isolated quality metrics. Because bounded metrics (like 1 to 5 stars) often form a continuous density cloud, the algorithm mathematically “slices” the data.

Actionable Business Segmentation

To make these insights highly actionable for Procurement and Quality Assurance teams, we pivot to a Business Logic Matrix. This definitively groups products based on acceptable risk thresholds.

Strategic Summary & Next Steps

Based on our strictly isolated quality segmentation, here is the breakdown of our active product catalog:

Recommended Actions:

• Premium Quality: Allocate higher marketing spend and feature these products prominently on the front page.
• Mixed/Inconsistent: Investigate supplier consistency. Audit recent reviews for specific common complaints.
• High Defect Risk: Immediately pause purchasing or delist until a supplier quality audit is completed.

UPCOMING INITIATIVE: Predictive Risk Modeling

Moving from Reactive to Proactive While our current clustering model tells us which products have historically failed, our next initiative uses Supervised Machine Learning to predict low-quality customer experiences (1-star or 2-star reviews) before they even happen.

Our Modeling Strategy features 3 Tiers: 1. Model A (Pre-Shipment Risk): A Logistic Regression model estimating risk purely based on listing attributes (price, weight, freight cost, description length, and photos). 2. Model B (Full Customer Experience): Adds real-world delivery outcomes (total delivery time, delay in days) to assess the complete journey. 3. Model C (Non-linear Challenger): A powerful Random Forest algorithm designed to capture complex interactions between all variables.

UPCOMING INITIATIVE: Evaluating The Models

By training on 80% of our historical data and evaluating on a 20% holdout set, we anticipate strong predictive power.

UPCOMING INITIATIVE: Business Value & Application

The Random Forest and Logistic Regression models will output feature importance and odds ratios, answering: What exactly drives a negative review?

How we will use this: Every new order will be scored in real-time. Orders flagged as high-risk will trigger automated workflows—such as proactive customer service reach-outs or expedited shipping—saving the customer relationship before it sours.