The Pulse of Delivery

In the world of quick commerce, a minute isn’t just a measure of time—it’s a measure of trust. When a customer taps “Order,” they aren’t just buying groceries; they are buying a promise. But what happens when that promise is broken?

For my latest project, “Customer Segmentation Analysis”, I didn’t just want to group customers by how much they bought. I wanted to understand why they bought—and more importantly, why they might stop. By analyzing transactional data from a quick-commerce platform, I set out to answer a simple question: Does delivery reliability shape the type of customer you attract?

The answer, as the data revealed, was a resounding yes.

---

The Approach: Beyond RFM

Traditional customer segmentation often relies on RFM analysis (Recency, Frequency, Monetary value). While powerful, this approach often ignores the experience of the customer. I hypothesized that operational metrics—specifically delivery delays—were just as critical as spending habits.

The Toolkit

• Python for data processing.
• Pandas & NumPy for rigorous data cleaning and manipulation.
• Scikit-Learn for clustering (K-Means) and dimensionality reduction (PCA).
• Seaborn & Matplotlib for bringing the data to life.

Data Cleaning: The “Messy” Reality

Real-world data is rarely clean. My dataset contained over 50% missing values in some columns, significant outliers in order totals, and negative delivery times (time travel isn’t a feature yet!).

Key steps in the pipeline included:

1. Imputation: Using median/mode strategies for missing data to preserve distributions.
2. Outlier Removal: Capping order_total and removing extreme delivery_delay_minutes using the Interquartile Range (IQR) method.
3. Feature Engineering: I created a custom feature set that combined transactional behavior with operational reality:
   • avg_order_total: How much they usually spend.
   • total_spent: Lifetime value (log-transformed to handle skew).
   • num_orders: Frequency of purchase.
   • avg_delay: The average operational friction experienced by the customer.

---

The Discovery: Two Worlds of Customers

After preprocessing and scaling the data, I used K-Means Clustering to segment the user base. To find the optimal number of clusters (k), I evaluated several metrics:

• Silhouette Score: Measures how distinct the clusters are.
• Davies-Bouldin Index: Measures cluster compactness.
• Calinski-Harabasz Score: Measures variance ratio.

All three metrics pointed to an optimal k of 2. This might seem simple, but the simplicity revealed a stark polarization in the customer base.

Cluster 0: The “On-Time High-Value” Segment

These are the ideal customers. They order frequently, spend significantly more per order, and—crucially—they experience very few delays. Reliability breeds loyalty.

Cluster 1: The “Delay-Affected Low-Value” Segment

This group tells a cautionary tale. They order infrequently and spend less. But the defining characteristic? They suffer from consistent, significant delivery delays.

(Place your PCA Scatterplot here to visually show the distinct separation between these two groups)

---

Features that Matter

I used Principal Component Analysis (PCA) to visualize these high-dimensional relationships in 2D space, confirming the separation. But looking at the feature importance gave the actionable insight:

• Order Frequency & Total Spend were strongly positively correlated with On-Time Deliveries.
• Delivery Delay was the primary wedge driving customers into the lower-value segment.

It suggests a feedback loop: A customer experiences a delay -> they order less/spend less -> they become “low value” (or churn entirely).

---

Conclusion: Operational Excellence is Marketing

This project went beyond a simple coding exercise. It demonstrated that in quick commerce, operations is marketing. You cannot segment high-value customers solely by marketing to them; you have to deliver to them—literally.

The “Delay-Affected” cluster represents a massive opportunity. If the business can fix its operational bottlenecks for these specific users, there is a clear path to migrating them into the “High-Value” tier.

Check out the full analysis and code in the main project notebook.