Data mining · association · RFM

data-mining-retail

A reproducible fictional retail pipeline ranks cross-sell rules by lift and turns Recency, Frequency, and Monetary behavior into explainable segments.

Public · synthetic demo

The problem

Popular items can make weak pairings look strong.

Confidence alone rewards common products. Lift asks whether items co-occur more than chance, while RFM separates valuable but quiet customers from generic top spenders.

Input

A fixed fictional transaction history

3,000 transactions · 300 customers · 16 items
customer_id, transaction_date, item, price
Planted patterns include milk+sugar→coffee
and butter+eggs→bread.

The money shot

Measured associations and actionable customer states

Committed top association rules by lift chart — Rules ranked by excess co-occurrence over chance.

Committed customer segments chart — All 300 fictional customers assigned to one RFM segment.

Antecedent → consequent	Support	Confidence	Lift
milk + sugar → coffee	0.0803	0.7303	3.926
butter + eggs → bread	0.0580	0.5421	2.978
apples + pasta_sauce → pasta	0.0283	0.5120	2.839
bacon + pasta → pasta_sauce	0.0220	0.7952	2.764

Segment	Customers	Avg recency	Avg frequency	Avg monetary
Champions	56	3.5	12.5	254.1
At-risk	49	50.5	12.7	137.6
Lost	52	79.8	7.1	65.1

At-risk customers bought frequently but have gone quiet—the win-back group a top-spender list misses.

How it's verified

Explainable math, not a black box.

Apriori support, confidence, and lift are implemented directly; RFM scores use deterministic quantiles. Tests verify metric math on toy baskets, planted-rule recovery, valid score ranges, complete segmentation, and that Champions are more recent and valuable than Lost.

Honest limitations

All transactions, rules, and segments are fictional and illustrate the method.
Support and lift thresholds must be tuned to catalogue size and business cost.
Large catalogues require lower support carefully and a capped itemset length to control Apriori cost.
Association is not causation; test merchandising actions before rollout.