Part 2 · How AI Actually Works
Finding Patterns Without Labels: Clustering
Classification sorts new items into categories you already defined. But what happens when you do not know the categories yet? Hand a child a box of mixed toys and say nothing. Within minutes the cars sit in one pile, the blocks in another. Nobody named the groups. The child found them.
Why This Matters
Some of the most valuable patterns in your work are ones nobody has labelled yet, the natural groupings hiding in your customers, your documents, your transactions. Knowing when a machine can surface those groups is part of understanding what AI can and cannot do. By the end of this lesson you will be able to describe how clustering differs from classification, explain when clustering is the right approach, and identify the tasks where discovering hidden groups would be valuable.
Groups You Did Not Define
Classification learns from examples you have already labelled: you define the categories, show the machine thousands of sorted examples, and it places new items into those known buckets. Clustering starts with no labels: you hand it the raw data, and it groups items by how similar they are, forming clusters you never defined. Because you supplied no answers, this is unsupervised learning, the other family from classification's supervised kind.
| Classification | Clustering |
|---|---|
| You define the categories Known in advance | No categories given You do not know them yet |
| Learns from labelled examples Sorted data to train on | Groups by similarity Near points sit together |
| Sorts new items into known buckets Places, does not discover | Discovers structure you never defined Surfaces, then you interpret |
The mechanism is simpler than it sounds. Clustering measures how close data points sit to one another and puts the near ones together. So it is the right approach whenever you do not know the groups in advance and want the data to reveal them: segmenting customers by behaviour, grouping support tickets by theme, or flagging transactions that fit no normal pattern.
Consider a retailer who assumes two kinds of customers: regulars and occasional shoppers. Feed a year of purchase data to a clustering model with no labels, and five distinct groups can emerge: weekday lunch regulars, weekend bulk buyers, promotion-only shoppers, and more. Patterns the team suspected but could never prove.
One caution. The machine finds the groups; it cannot tell you what they mean or whether they matter, and the clusters only reflect the data you fed in. The discovery is the machine's. The interpretation stays human.
Key Takeaway
The discovery is the machine's; the interpretation stays human.
Check your understanding
Self-check only. Nothing is scored and nothing is recorded; open an answer to compare it with your own.
A retail chain has transaction data from 50,000 customers but has never formally segmented them. The marketing director says: 'We just treat everyone the same, we send the same promotions to all customers.' A data analyst suggests using clustering. What would clustering do in this situation?
- Sort customers into pre-defined segments like 'premium,' 'regular,' and 'budget.'
- Predict how much each customer will spend next month.
- Analyse purchase patterns and discover natural customer groups that the team has not formally defined, then the marketing team can tailor promotions to each group.
- Replace the marketing team's judgement with an algorithm that decides all promotional strategy.
Show the answer
Analyse purchase patterns and discover natural customer groups that the team has not formally defined, then the marketing team can tailor promotions to each group.
Your colleague says: 'We already sort customer tickets into five categories using AI classification. Why would we also need clustering?' Based on what you learned, what is the best response?
- You do not need both, classification and clustering do the same thing.
- Clustering could reveal that the five categories do not capture reality, maybe there are actually seven natural groups, or maybe two of the five categories should be merged because the underlying issues are the same.
- Clustering is only useful when you have no categories at all.
- Clustering should replace classification entirely, it is more advanced.
Show the answer
Clustering could reveal that the five categories do not capture reality, maybe there are actually seven natural groups, or maybe two of the five categories should be merged because the underlying issues are the same.
Saved in this browser only. Carry progress between devices with the resume link on the course page.