Back to evidence library

04 / Paris-Dauphine · 2025

Live demo · Public sourcedata science

Customer segmentation decision dashboard

From clustering diagnostics to an interface a marketing team can actually explore.

The work combines cleaning, feature engineering, K-means/CAH/GMM comparison and an interactive dashboard. The emphasis is not only on cluster labels but on the behaviors a stakeholder can inspect before designing a campaign.

Segmentation project visual

Evidence register

2,240

customers

Behavioral and demographic observations

3

methods

K-means, hierarchical clustering and GMM

4

segments

Selected with elbow and silhouette diagnostics

01 / Problem

A single marketing message ignores meaningful differences in purchasing behavior, engagement, recency and channel preference.

02 / Approach

After cleaning and feature engineering, three clustering families are compared with internal diagnostics before the selected partition is translated into stakeholder-readable profiles.

03 / Outcome

The deployed Shiny dashboard supports filtering, PCA views, cluster sizes, radar profiles and re-running the selected clustering configuration.

How the evidence is produced.

Led data preparation, method comparison, interpretation and full Shiny dashboard implementation.

  1. 01Raw customer table → cleaning and features
  2. 02Scaled matrix → three clustering candidates
  3. 03Internal diagnostics → selected partition
  4. 04Segment profiles → interactive Shiny dashboard

Validation scope

Internal validation uses elbow, silhouette and Davies-Bouldin diagnostics. The current silhouette is approximately 0.35, indicating useful but overlapping groups.

Known limitation

The segments have not yet been validated through campaign uplift, temporal stability or out-of-sample assignment. They support exploration, not causal targeting claims.

What is inspectable

  • The interface exposes both diagnostic and stakeholder views.
  • Cluster descriptions connect behavior, recency, spending and channels.
  • The live application is independently accessible without local setup.

Next proof to add

  1. 01Lock the R environment and add an automated Shiny smoke test.
  2. 02Measure cluster stability under resampling and across time windows.
  3. 03Validate campaign value through an uplift or controlled-experiment design.

Main stack

RShinyK-meansCAHGMMPlotly