Who are your customers, really?

What is RFM?

RFM is a customer-scoring framework used in CRM and marketing to rank customers on three behavioral dimensions:

• Recency — how recently the customer last purchased. Lower is better.
• Frequency — how often they purchase. Higher is better.
• Monetary — how much they spend overall. Higher is better.

Customers with strong scores across all three are your most valuable; weak scores across all three are churn risks. The three-way split also separates customer types that pure revenue ranking would conflate — e.g. a high-spend customer with stale recency behaves very differently from a high-spend customer with recent activity.

How RFM maps to the dataset

Total_Purchases is the sum of all Num*Purchases channels (deals + web + catalog + store), and Total_Spent is the sum of all Mnt* product categories — both engineered during the cleaning phase.

From clusters to a score

Why K-Means? K-Means is an unsupervised algorithm that partitions data into k groups by minimizing within-cluster variance. It's a natural fit for RFM because each dimension has clear “tiers” of customer behavior that aren't pre-labeled.

Why three 1D models, not one 3D model? This pipeline runs K-Means once per RFM dimension — three separate models on Total_Purchases, Total_Spent, and Recency. This avoids two pitfalls of joint 3D clustering:

1. The three dimensions live on wildly different scales (days vs. dollars vs. counts), so joint clustering would require careful normalization.
2. Each dimension has its own natural breakpoints — running K-Means independently respects them.

Hyperparameter search. Each 1D model is tuned with RandomizedSearchCV over n_clusters ∈ [2..4], init ∈ {'k-means++', 'random'}, n_init ∈ [10..50], max_iter ∈ [100..700]. The cluster cap of 4 is deliberate: too many clusters per dimension makes the final segments hard to translate into marketing actions. All three searches converged on k = 4.

From cluster label to score. The cluster ID itself is meaningless (K-Means assigns IDs arbitrarily), so each component's clusters are re-ranked by their average value:

• Frequency_Score and Monetary_Score — clusters sorted ascending by average → scores 0→3.
• Recency_Score — clusters sorted descending by average → scores 0→3 (recent = high score).

The final RFM score. The three scores sum into a single RMF_Score from 0 to 9, which is bucketed into five business segments: