Tae Hyun Kim (Lowell)
Personalization

Customer Segmentation

1 min read #targeting#segmentation

Definition

Customer Segmentation is the unsupervised task of partitioning customers into a finite set of segments by similarity in behavior, value, and preference. A common recipe is latent-factor decomposition followed by clustering: behavioral features → NMF (non-negative, parts-based decomposition) → factor scores → K-Means → segments.

Intuitive Understanding

“You cannot manage thousands of people individually, so you reduce them into groups that behave similarly and differentiate your strategy accordingly.” The key is that segments must be actionable and reproducible.

Role in Targeting

A segment is not an end in itself but an HTE moderator — because uplift differs across segments, the segment becomes an axis for interpreting why heterogeneous effects arise. (If User Profiling is an individual-level representation, a segment is the summary of its cluster.)

Quality and Stability Diagnostics

  • Internal quality: Silhouette, Davies-Bouldin, Calinski-Harabasz (is the current assignment good?)
  • Reproducibility: Bootstrap Stability + ARI (does this assignment reproduce?)
  • NMF vs PCA: preferred for behavioral data because of its non-negative, parts-based interpretability.

Project Application

Dunnhumby: 5 latent factors explain 92.44% of behavioral variance, yielding 7 segments (Bootstrap ARI 0.767 ± 0.113). The high-value segment, 45% of customers, contributes 68.4% of revenue. (project canonical)

  • Targeting Overview ← hub
  • NMF — latent-factor extraction
  • User Profiling — individual-level preference representation (the atomic unit of a segment)
  • HTE — segments are moderators of effect heterogeneity

References

  • MOC-Targeting

Local graph