Clustering Is Where Data Starts Talking 🤖📊

Most people in ML jump straight into future prediction.

Fraud. Churn. Next best offer. I did jump to regression and classification first too. But over time, I realized something foundational:

If you don’t understand the structure of your data, your predictions are fragile. Clustering fixes that.

Before predicting what will happen, you need to understand what already exists.

What Clustering Really Is 🔍

Clustering is “unsupervised” learning. No labels. No predefined truth.

You take raw data and ask: What looks similar?

That’s it. No “good customer” label. No “risky customer” tag. Just patterns emerging.

Clustering is not about accuracy. It is about clarity about the information at hand. You use it when:

• You want to simplify complexity
• You suspect hidden patterns
• You don’t know your customer segments
• You want to pinpoint anomalies…

It turns chaos into structure and structure drives decisions.

Where You Already Use It (Even If You Don’t Know) 🧠

Clustering sits behind:

• Banking customer segmentation
• Product recommendations at an e-commerce site
• Risk grouping in credit lending
• Behavioral analysis of software users
• Marketing strategy of a big retailer
• RAG retrieving chunks of text based on similarity

Have you ever seen a segmentation slide in a boardroom?

Clustering was behind it…

The Part Nobody Talks About ⚙️

People argue about algorithms. Hierarchical clustering, K-Means/K-Medoids, or DBSCAN. These are important but improving them is…

Wrong focus!

The real game is the richness of your feature space. Clustering does not understand customers. It understands vectors. You define the dimensions. You define reality.

Now this is changing fast.

With LLMs and VLMs, feature engineering is no longer limited to structured data. You can turn almost anything into vectors:

Text. Conversations. Documents. Voice. Images. User behavior…

This expands the space dramatically. You are no longer clustering only transactions or demographics. You are clustering intent, context and meaning. The bottleneck is shifting.

From “how do I cluster?” To “what representation of reality do I choose?”

That is a much bigger decision. Think about the MR. Fusion device from Back to the Future ⚡. You throw in a banana peel, a beer can, random scraps. Out comes usable energy.

That is what LLMs and VLMs did to feature engineering.

Before, you were limited to structured tables, clean numeric fields, predefined categories.

Now you can take customer emails, call center transcripts, product images, reviews, clickstreams.

Turn all of them into embeddings. Put them into the same vector space. And suddenly you are not clustering data types anymore.

You are clustering meaning.

Just like MR. Fusion turned random inputs into energy, embeddings turn unstructured chaos into structured signal. The difference is:

Mr. Fusion hides the complexity. Here, you still need to decide what to feed in. That decision defines the outcome.

A Real Project for a Tier 1 Bank 🏦

I led segmentation projects in banking, did it for telecoms, insurance, retail and logistics as well…

The biggest one was in a Tier 1 bank with 20 million customers. We built a feature space with 41 variables. Not random. Very intentional.

• Assets, asset potential
• Debts, credit potential
• Product ownership
• Channel behavior
• Socio-demographics

The goal was clear:

Move away from “what the customer has”, it’s a snapshot of today!

Move toward “what the customer needs vs can afford”, it’s about the near future and investing in it.**

That shift changed the outcome. Our version of customer segmentation was built on:

• Value dimensions like assets, investment depth and investment variety. Also credit products, debt levels and usage richness.
• Behavior dimensions included product details and physical/digital channel usage
• Socio-demographics like age, occupation and income signals

We ended up with three macro groups:

• Asset potential
• Credit potential
• Indifferent

And dozens of actionable sub-segments. Not only clusters. Strategies to act on. We got:

• High asset, low credit usage → investment opportunity
• High debt, high potential → lending expansion
• Low engagement → low priority or activation campaigns

Each segment linked to revenue, risk and product strategy. That is when clustering becomes real.

Why Most Clustering Work Fails ❌

I see this all the time:

• Picking random features
• No business hypothesis behind
• Overfitting clusters to metrics
• No linkage of clusters to decisions

Result? Nice charts. Zero impact.

What Works Instead ✅

If you want real value:

1. Start with a business question, not an algorithm
2. Design your feature space carefully, this is 80% of the work
3. Choose algorithms based on structure, not popularity
4. Validate with business metrics: Revenue. Risk. Behavior
5. Turn clusters into actions: Campaigns. Pricing. Offers

If no decision changes, the work has no value.

From Consulting to Vibe Coding 💻⚡

While thinking about these projects lately, I wanted to explain the intricacies of clustering to my colleagues and people I teach about machine learning and AI. This is exactly why I built a clustering sandbox.

This is my second vibe coding experiment. Simple idea: Make clustering visible. You can:

• Generate synthetic datasets
• Run 5 algorithms side by side
• Compare metrics instantly
• See failure cases clearly
• Replay historic datasets and parameters too…

Because there is no “best algorithm”. Only better understanding.

[embed]GitHub - aatabarezz/clustering-sandbox: Interactive clustering sandbox - K-Means, DBSCAN, HDBSCAN… Interactive clustering sandbox - K-Means, DBSCAN, HDBSCAN, Hierarchical & Mean Shift with auto-estimators, DBCV, run…github.com

Final Thought 🎯

Prediction is powerful. But structure is fundamental. Clustering gives you that structure. And in my experience:

This is where data stops being noise and starts becoming strategy.

All the best

Altan

P.s.

Some great reading and python code on different clustering algorithms their pros and cons;

[embed]Comparing The-State-of-The-Art Clustering Algorithms Let’s generate complex data and try different clustering algorithmsmedium.com

Also here is a great Medium piece on how LLM’s improve customer segmentation big time!

[embed]Mastering Customer Segmentation with LLM Unlock advanced customer segmentation techniques using LLMs, and improve your clustering models with advanced…medium.com