Discover hidden groups in your data with Clustering: Get business-friendly, reusable labels

This article focuses on Clustering models in Einstein Studio’s Model Builder, designed to simplify identification of data-driven clusters easily. No more guessing sub-segments or cohorts in your data with complex IF-THEN-ELSE conditions. Let the model discover clusters and also generate business-friendly labels

This blog is the first of a 3-part series on Clustering. 

This explains Einstein Studio’s “Supervised” Clustering with the use case of Customer segmentation, one of the most common CDP use cases, but is scalable to any kind of cohorting on products, accounts, customers, assets or combinations. 

What you will learn from this blog: 

• Realize the outcomes of Clustering with 2 industry use cases. 
• Understand what Einstein Studio clustering offers and its differentiators
• Build the model in Einstein Studio.
• Get a cheatsheet on when to use Clustering vs. Multi-class Classification

Next parts will talk about the traditional Unsupervised clustering and Semantic Clustering.

Let’s get started on the blog. 

Customer segmentation via Clustering: 

Customer segmentation is one of the most fundamental jobs of a marketer,  knowing who your customers are, grouping them into cohorts, and running campaigns that actually resonate. Today, most teams do this by hand-writing filters in segments, which works until the customer base grows or the profiles get subtle. 

Let’s understand how Clustering can enhance this process and grow revenue with 2 examples. 

🛒 Retail: Finding the customers your tiers missed or grouped together

A mid-sized retail brand has 500,000 customers sorted into Gold, Silver, and Bronze tiers using some RFM-based rules. The rules are clear, the campaigns are running, and every customer who matches a tier gets a treatment. 

But two things quietly go wrong: 

• Hidden across tiers: Distinct behaviors / micro-segments get averaged into a single tier, and 
• Unclassified customers: Customers who don’t belong to any rules, slip out of every campaign.

Imagine they now run a clustering model on the same behavioral data containing RFM i.e. recency, frequency, monetary value with no labels, no rules

Notice what changed. 

• BEFORE: There were 3 hard-coded rules to split as tiers 
• AFTER –  With Clustering, Marketers now see multiple clusters; either split from a single tier, or unknown clusters identified. They can now run specific marketing campaigns as seen in the images: 
  • New sub-clusters identified:
    • Sporadic big spenders who show up occasionally with large baskets, and 
    • Steady low-basket loyalists who buy small and often.
    • The tier had been averaging them into one “steady middle”.  
  • A new At-Risk cohort that cuts across Gold, Silver, and Bronze identified: earlier you only had threshold based, not spending pattern based. 
  • And the Unclassified customers who never matched a rule now form an Emerging Explorers cohort worth a welcome journey. 

We’ll cover how these labels were generated in a subsequent section.

Now another industry but with a customer profiling angle.

🏦 Loans: Finding the applicant profiles your rules couldn’t see

A regional bank processes 200,000 loan applications a year. Marketing runs a handful of campaigns ,  an offer here, a generic campaign there. Response rates are okay, but the CMO wants sharper Applicant Profiles ,  the lending-equivalent of Ideal Customer Profiles (ICPs),  so marketing can match the right offer and channel to each profile.

They now run a clustering model on the applicant data they already have with no pre-defined labels.

The model surfaces the natural applicant profiles hiding in the data, along with the top factors that shape each one. Marketing can now possibly look at different campaign, engagement strategies for each profile. 

1. Older Medium-Income Borrowers: defined by age and income: The bank can handle them with top-up and refinance offers and engage over branch and phone
2. Young High-Income Couples: defined by age, and marital status.  The Bank can give first home-loan offers and engage over mobile and referral channels
3. Affluent High Loan Seekers: they can get high-ticket home-loan offers, preferential rates, and personalized treatment via relationship manager. 

Same data, three targeted plays,  no one-size-fits-all blast.

To summarize, Clustering goes beyond what rules-based criteria can go and uncover / discover profiles and get you to better marketing strategies.

👥 Clustering vs. Multi-Class Classification, A Quick Cheat Sheet

A common question arises now: isn’t this just multi-class classification? Not quite ,  here’s the difference

Multi-class classification  is a model trained on labelled data 

• to predict from a list of existing categories like Channel optimization or 
• to predict from a list of categories that you define manually like the Gold, Silver and Bronze tiers. 

If you use it for use cases like customer segmentation, you are replicating the same rules but the labels prepared via Transforms / CIs, get into your model and have the exact setup. 

Bottomline: Use clustering when the question is: “What groups naturally exist here that I don’t know about?”. Let it discover these groups.

🔍 How is Einstein Studio’s Clustering better than the traditional Unsupervised k-means models?

The Challenges with Traditional Unsupervised k-means models

1. Unstable clusters across runs: With k-means clustering (or any other clustering algo) every time you run the model on new data, the clusters will mostly shift. A cluster labelled “Cluster 3” last month might represent completely different customers this month. That instability breaks your downstream processes like your Segments, Flows, Tableau reports.
2. No Cluster labels: Kmeans only provides Cluster ids, which are not understandable by business users. 

What Makes Einstein Studio Clustering Different

1. Consistent, Stable labels across every run (Key Differentiator from AutoML tools ⭐). Einstein Studio solves this with a “supervised approach”. You train a model,  lock the clusters for reuse.  Any new record gets mapped to the same set of clusters, avoiding downstream breakage.

2. Business user friendly labels to understand clusters better (Key Differentiator from AutoML tools ⭐). Standard k-means gives you just the cluster IDs: Cluster 0, Cluster 1, Cluster 2 are not interpretable or useful to a marketer.
   Einstein Studio allows users to automatically generate human-readable, business-meaningful names for each cluster. You can even provide custom instructions on your industry or brand vocabulary.

3. Transparent model metrics: Simple metrics for business users on whether the model is ready to deploy, alongside advanced metrics for any data scientists who want to go deeper. No black boxes.

4. Cluster summary: The model provides you with a list view to interpret the distribution, and a scatter plot view to visualize your clusters in a 2D space. Drill down to each cluster to see sample records or the value distribution.
   Cluster summary with the distribution chart against each of the IDs (and the labels created using the Generate labels button).

5. SHAP / Interpretability at 3 levels:
   1. Model level SHAP available in Training metrics, and Overview page
   2. Cluster level SHAP to understand driving factors for each cluster. Notice how for each of the clusters, the features and its importance are different.
   3. Local / Prediction level SHAP: This explains the prediction itself and shows which factors contributed to this prediction and by how much. 

⚙️ How to get started on Clustering models in 3 steps: 

Lets build a Clustering model using RFM based segmentation and answer a complex question: Who are my most valuable customers right now based on their behavior?

⚙️Data prep: 

If you have Customer data, and Sales order data, you can get started. 

• On Sales Order DMO, create metrics on RFM using given logic

Feature	Description	Transform logic
Customer_ID	Identifier	Use as is
Recency (R)	Days since last purchase (Lower = better)	DATEDIFF(CURRENT_DATE, MAX(order_date)) grouped by Customer_ID
Frequency (F)	Total number of purchases over a time window, say 30 days(Higher = better)	COUNT(order_id) grouped by Customer_ID
Monetary value (M)	Total spend over a time window, say 30 days(Higher = better)	SUM(order_amount)grouped by Customer_ID

• Join with the Customer data against Customer_ID
• On Customer DMO, select a few demographic attributes – age, gender, other business-specific attributes 
• Save this as a DMO “Customer segmentation training DMO” to use in Einstein Studio

🚆Train the model

Go to Einstein Studio → Select Cluster in the left pane and Click on Add Clustering Model.

This is pretty similar to building a predictive model, except that there is no choosing the goal column.

• Select the Training DMO in the model builder.
• Choose the right variables. The clusters are feature-dependent. So choose appropriate variables which you think will impact the clusters.

• Choose the right algorithm – kmeans when you want a specific number of clusters (k) or use auto-k to let the model decide. Use HDBScan when there are outliers. 

• Save and Train.
• View Training metrics and Activate the model.

▶️ Watch the step by step instructions in this demo video.

🏷️ Generating labels for your clusters

• By Default, the labels are listed as Cluster 0, Cluster 1, etc which is not easy to interpret, understandable or usable even with the Data preview of the clusters.

• Click on Generate Labels and provide instructions related to your industry / use case.

• View the new labels generated using the context given and based on the cluster’s metadata and sample data points. Regenerate to tweak further or manually edit the names or save the names.

🚀Putting the trained cluster model to work

Discovering clusters is only half the value. The other half is what you do with them. Use the model across inference channels from the Integrations tab.

1. Use in batch transforms or predict jobs to write the data into a DMO.

2. Use the cluster outputs from Transforms in Segments. 

Filter for 

• Cluster Label = Top Loyal Spenders to build a precision audience of your highest-value shoppers
• Add this as a complementary rule to your existing segment or keep a standalone Clustering based segment.

3. Use the clustering models in Flows

Get cluster labels for each record in Flow and create downstream automation for each of the clusters.
For example, 

• Route At-Risk customers into a time-boxed win-back journey regardless of the tier they wear today, or 
• Send Emerging Explorers into a welcome journey with a first-purchase incentive

Scaling Clustering use cases across your business

Customer segmentation is the headline use case, but the same capability can find anomalous supplier patterns in an operations dataset, or cluster on Cases, or Vendors / Suppliers based on their demographic data and behaviors.

Conclusion

Clustering is how you stop segmenting the customers you already know and start discovering the ones you don’t.

Rules-based segments will always have their place – they’re transparent, easy to explain, and easy to audit. But they only find what you already thought to look for.

Clustering reads across features you provide on recency, spend, trajectory, and behavior together and surfaces the cohorts your filters averaged into one tier or missed entirely.

Generate labels and power segments, Flows or any downstream better.

Note: Special thanks to Bobby Brill for the review.
You can read all our previous blogs here: Model Builder – salesforceblogger.com. This blog’s content is also available as a How-to demo video on Clustering models and a whole lot of other Demo videos across the ML lifecycle are available in our Einstein Studio Demo playlist.

Written by