Customer Segmentation

Advanced customer segmentation solution to tailor marketing strategies to specific user groups.

Feature-Rich Customer Segmentation

Executive Summary

One-size-fits-all marketing is dead. To stay relevant, businesses must understand the diverse personas within their customer base. This project delivered a comprehensive Customer Segmentation engine using unsupervised learning (Clustering). It identified 6 distinct customer personas (e.g., “Bargain Hunters,” “Loyal Enthusiasts,” “Dormant High-Value”), enabling the marketing team to tailor messaging, offers, and product recommendations for each group, resulting in a 14% increase in email engagement.

Problem Statement

The business relied on basic heuristic segmentation (e.g., “Bought in last 30 days”). This failed to capture behavioral nuances—a customer buying a single expensive item once is very different from one buying cheap items weekly, yet they might look similar in total spend. We needed a multi-dimensional segmentation approach.

Methodology

1. Feature Selection (RFM+)

  • RFM: Recency, Frequency, Monetary Value.
  • Behavioral: Average time on site, preferred categories, discount sensitivity (ratio of purchases made on sale), return rate.
  • Demographic: Location tier, age group.

2. Preprocessing

  • Scaling: Standardized features using StandardScaler (Z-score normalization) to ensure features with large ranges (e.g., Revenue) didn’t dominate features with small ranges (e.g., Frequency).
  • transformation: Applied log transformation to skewed distributions (like Spend).

3. Clustering Approach

  • K-Means: Selected for its efficiency and interpretability.
  • Elbow Method: Used to determine the optimal number of clusters (k=6).
  • Silhouette Score: Validated cluster quality/separation.

Implementation Details

The segments are recalculated weekly to account for changing user behavior.

  • Pipeline: Python (Pandas + Scikit-Learn) script running on AWS Lambda (triggered by EventBridge).
  • Output: A segment_id and segment_name tag is applied to each user profile in the Data Warehouse (Snowflake).
  • Integration: These tags are synced to the Email Service Provider (ESP) and Push Notification tools appropriately.

The 6 Personas identified:

  1. Champions: High spend, high frequency, recent. (Strategy: VIP rewards, early access).
  2. Loyal Potential: High frequency, low spend. (Strategy: Upsell/Cross-sell).
  3. Big Spenders: Low frequency, high spend. (Strategy: Nurture with premium content).
  4. Promiscuous: Only buy on deep discount. (Strategy: Flash sales, clearance).
  5. At Risk: Previously good customers, high recency. (Strategy: Win-back campaigns).
  6. Hibernate: Low value, haven’t visited in long time. (Strategy: Low-cost re-engagement).

Challenges & Solutions

  • Challenge: Interpretability. “What does Cluster 3 mean?”

  • Solution: Created “Snake Plots” and “Relative Importance Heatmaps” to visualize how each cluster differs from the population average on every feature.

  • Challenge: Stability. Customers “jumping” between segments too frequently.
  • Solution: Implemented a smoothing logic where a user must exhibit behaviors of a new segment for 2 consecutive weeks before being moved, preventing marketing whiplash.

Results and Impact

  • Engagement: Customized subject lines for “Bargain Hunters” vs “Champions” led to a 14% increase in Open Rates.
  • Revenue: “At Risk” win-back campaigns recovered $50k/month in potentially lost revenue.
  • Strategy: This segmentation is now the “common language” used across Product, Marketing, and Sales teams to discuss user groups.

Future Work

  • Micro-Segmentation: Breaking down these 6 macro-segments into smaller niches (e.g., “At Risk” -> “At Risk - High Value” vs “At Risk - Low Value”).
  • Persona Evolution: Tracking how users move between segments over their lifecycle (Markov Chains) to predict LTV trajectories.