The question of whether to fold richer dimensions into an RFM clustering, or keep them separate, deserves a direct answer: don't mix them in one pass. Treating behavioral variables like quarterly purchase ratios, return rates, or channel preferences as just more columns for a single k-means run dilutes the clarity that made your original RFM clusters useful in the first place. RFM is a segmentation of value and engagement. Adding navigation data or payment methods introduces a different kind of signal, one that answers a different question. When you blend them, you force one algorithm to optimize for two competing objectives, and the result is often a compromise that satisfies neither.
A more practical path is to use your RFM clusters as the foundation, then layer additional dimensions on top. Run a second clustering within each RFM segment using the new variables, returns, channel mix, ratio of web to store purchases. This keeps the core value segments intact while revealing meaningful sub-patterns inside them. For example, your high-value segment might split into "consistent online buyers who rarely return" versus "omnichannel shoppers with higher return rates." That distinction is actionable. It tells you where to invest in retention, where to adjust shipping policies, and how to tailor messaging. If you had mixed everything into one model, you might never see that split clearly.
What about the risk of correlated variables, like ratio Q2 to Q1 and ratio Q3 to Q2? Yes, it's a real concern. Highly correlated inputs can over-weight a single underlying trend, effectively doubling its influence without adding new information. The fix is not to avoid such variables but to check their correlation before clustering and consider dropping or combining the most redundant ones. You can also use a simple validation step: run the clustering with and without the extra variables, then compare cluster stability and how distinctly the segments differ on business metrics like repurchase rate or margin. If adding a variable doesn't improve separation or actionability, leave it out. That's the test.
So the procedure is clear. Start with RFM as your anchor. Then, within each cluster, experiment with a second k-means using the additional dimensions, but only after checking correlations and validating that the new segments are stable and meaningful. Use silhouette scores or similar metrics to compare, but don't rely on them alone. Ask yourself: does this new split change what I would do next week? If the answer is no, it's not worth the complexity. And for web navigation data specifically, treat it as a separate layer entirely. It's a different behavioral context, and forcing it into the same model as transactional RFM risks noise. Keep it distinct, validate it on its own terms, and only integrate it when it clearly drives a different decision. That's how you turn richer data into sharper action, without losing the signal that made RFM valuable in the first place.