Ifrah Rashid

The project started with real-world retail transaction data — the kind that arrives inconsistent rather than analysis-ready. Before any modelling, the work was cleaning: resolving missing and malformed records, removing cancellations and invalid quantities, standardising customer and product identifiers, and checking that totals reconciled. Exploratory analysis then described the shape of the customer base: how purchase value distributes, how often customers return, and how heavily the data skews toward a small group of frequent buyers.

Raw transactions do not describe customers — they describe events. RFM converts an event log into a behavioural profile per customer, which is what the clustering algorithms actually need. Skewed monetary and frequency distributions were transformed and scaled so that distance-based algorithms would not be dominated by a handful of extreme spenders.

• R

  Recency

  How long since this customer last purchased?

• F

  Frequency

  How often do they come back?

• M

  Monetary

  How much value have they contributed?

Rather than assuming one algorithm was correct, four were fitted to the same feature space and treated as four competing answers to the same question. Each carries a different assumption about what a 'segment' is — and those assumptions, not the code, are what decide the result.

Unsupervised learning has no ground truth, so the comparison relied on internal validation indices read together rather than any single score. Each measures a different property of a partition, and a model that wins on one can lose on another — which is exactly why more than one was used.

• Higher is better

  Silhouette Score

  How well each point sits inside its own cluster versus the nearest other one.

• Lower is better

  Davies–Bouldin Index

  Average similarity between each cluster and its most similar neighbour.

• Higher is better

  Calinski–Harabasz Score

  Ratio of between-cluster dispersion to within-cluster dispersion.

No single index decided the result — the three were read together, and against whether the resulting groups made business sense.

A cluster label is not an insight. The final step was reading each group's RFM profile back into plain business language, so that a retention or marketing decision could actually follow from it.

• RETAINHigh value, still activeProtect the relationship — these customers carry disproportionate revenue.
• RE-ENGAGEPreviously active, now lapsingReach out before the gap becomes permanent churn.
• NURTURERecent, low frequencyBuild the habit — turn a first purchase into a second.
• REWARDFrequent and loyalRecognise loyalty; test advocacy and referral programmes.

Model output and insight are not the same object. The clusters were the halfway point; the segments above are what a retention or loyalty decision can be built on.

1. 01

   Comparing models taught me more than tuning one. The disagreements between algorithms were the most informative part of the project.

2. 02

   Feature engineering decided the outcome more than algorithm choice did — RFM and the transformations applied to it shaped every partition that followed.

3. 03

   Clustering quality has to be argued, not asserted. Reading three indices together, and against the business meaning, was more honest than optimising one number.

4. 04

   Unsupervised results are hypotheses about structure, not answers. Interpretation is where the data science actually happens.

5. 05

   Translating a cluster into a decision someone can act on is a separate skill from producing the cluster — and the one that makes the work useful.