FutureSkilz.
Practise · 05

Skill Shot · Strategy · working

Choose a learning approach from the answer you need

By the end you can: choose between supervised classification and clustering by checking whether the outcome labels are already known.

About 7 minutes with AryaPrerequisite: Draft: awaiting human review

Arya · AI teacher

Arya is waiting.

Scene 1 of 7 · 0:00 to 0:55

Check whether customer outcomes are known before choosing an approach.

What changes when the team already knows who left?

Key points

  • Request: identify customers who may defect
  • Available: customer characteristics and past outcomes
  • Known outcomes: left or did not leave
  • Alternative question: discover customer segments

Illustrative case, written for this lesson

Customer defection question

  • The team wants to know which customers will defect.
  • Customer characteristics are available.
  • Past outcomes show who left and who did not.
  • The case is illustrative, not an actual organisation.

TakeawayStart with the question the analysis must answer, not with a preferred algorithm.

Arya says (captions)

Imagine a team wants to know which customers will defect. It has customer characteristics and records showing which customers left and which did not. The team could ask for a prediction about a known outcome, or it could look for customer segments without a predetermined answer. Those are not the same request. If we blur them, we may choose a learning approach that answers a different question. In this lesson, we will use one quick decision: are the outcome labels already known, or are we trying to discover groups? We will follow the customer example throughout, then test the decision on a new segmentation question. The case is illustrative, not a report about a real organisation. Keep the distinction in mind: predicting a known kind of outcome is different from exploring what groups exist.

Arya is an AI and the voice is synthetic. This is a draft not yet reviewed by a person.

Ask Arya

Stuck, or want it said another way?

Arya is an AI. She answers only from this lesson, gives hints and not answers, and asks you to try it in your own words. Please do not share personal details.

Sign in and your questions are kept with your profile; as a visitor, nothing you ask is saved.

Check yourself

Three quick questions

1. A team has examples with known outcomes and wants to sort observations into discrete classes. Which approach fits?

2. A team wants to find customer segments and has no known right answer for the groups. What should it choose?

Take it to work

This week, take one proposed analysis request and write down its desired result, whether outcome labels exist, and the approach that follows from that check.

Try it, then ask Arya

Your practice note

Write a few sentences on how you would use this on a real problem from your work: what happened, what you did, and what you expect. Arya will read it against what this Skill Shot teaches. Please leave out names and personal details.

0 words · about 12 or more gives a useful reading

How useful was this Skill Shot?

One tap. Your rating tells us what to improve.

Remember

  • Start with the result the analysis must produce.
  • Known discrete outcomes point to supervised classification.
  • Unknown groups with no right answer point to clustering.
  • The label check guides method choice but does not establish model quality.

Watch out for

  • Choosing a method before stating the requested result.
  • Calling a known-outcome prediction a clustering request.
  • Assuming the label check establishes data quality or guarantees a useful model.

Limits: This check helps select an approach; it does not assess data quality, model performance, or whether a model is appropriate for a particular decision.

Where this fits

Skill set
Strategy
Capability module
Choose a machine-learning approach from the outcome labels available
Learning path
New joiners and working professionals

Part 3 of 12

Next

  • Compare prediction and clustering with a labelled datasetPractise identifying the target outcome and the role of labels before exploring a wider range of machine-learning methods.
Full transcript and sources

1. Check whether customer outcomes are known before choosing an approach. · 55s

Imagine a team wants to know which customers will defect. It has customer characteristics and records showing which customers left and which did not. The team could ask for a prediction about a known outcome, or it could look for customer segments without a predetermined answer. Those are not the same request. If we blur them, we may choose a learning approach that answers a different question. In this lesson, we will use one quick decision: are the outcome labels already known, or are we trying to discover groups? We will follow the customer example throughout, then test the decision on a new segmentation question. The case is illustrative, not a report about a real organisation. Keep the distinction in mind: predicting a known kind of outcome is different from exploring what groups exist.

2. Choose the approach by checking for known outcome labels. · 55s

By the end, you should be able to choose between supervised classification and clustering for a customer-analysis request. You will do this by naming the result first, then checking whether examples already carry the correct outcome. If the request is to sort observations into known, discrete classes, and those outcomes are available as labels, supervised classification fits the question. If the request is to find customer segments and there is no right answer already supplied, clustering is the approach described in the notes. This is a bounded decision, not a promise that a model will work well. You will use the same check on the defection case and then apply it to a new segmentation problem. Success means you can state the requested result, say whether labels are known, and give a matching approach.

3. Use known labels for prediction and no labels to discover groups. · 58s

Here is the plain-language version. If examples come with the answer, a model can learn from those examples to predict that kind of answer for other observations. Think of a labelled example like an answer key: the outcome is already shown for the example. The proper term for learning from labelled examples is supervised learning. When the task is to sort observations into discrete classes, the notes describe classification. By contrast, if there is no right answer supplied and the aim is to find customer segments, clustering is an unsupervised learning technique. A non-example is calling a request to predict a known left-or-did-not-leave outcome clustering; that does not match the question. The distinction is about the target: known class to predict, or groups still to discover. It does not tell you whether the data or eventual model are good.

4. Apply a label-first check before selecting classification or clustering. · 57s

Use this label-first check. First, write down the result the analysis should produce: a known outcome, or groups that have not yet been identified. Second, ask whether the examples already show the correct outcome. Third, if the requested result is a known discrete class and examples are labelled, choose supervised classification. Fourth, if the aim is to discover customer segments and there is no right answer supplied, choose clustering. Keep the wording close to the request. “Which customers will defect?” asks for a known kind of outcome. “What customer segments exist?” asks for groups to discover. Do not choose an approach just because its name is familiar. This check is deliberately narrow: it helps match the question to an approach, but it does not assess the data, the model’s performance, or the consequences of acting on a prediction.

5. Choose supervised classification because the customer outcomes are labelled. · 57s

Return to the customer case. The requested result is to know which customers will defect. The available examples include customer characteristics and an outcome: the customer left, or the customer did not. The outcome is therefore known for those examples, and the requested result is a discrete class. Apply the check: name the result, confirm the labels, then choose supervised classification. The notes describe a process in which labelled training data are used to train an algorithm. Once in production, prediction data are used to make decisions, and new data can be used to improve the model over time. This process description does not change the initial choice: the request is still a prediction of a labelled outcome. It also does not prove that the model will make accurate predictions or that a particular business action should follow.

6. Choose clustering when the customer segments are not yet known. · 67s

Now make one decision on a new but related request. A team has customer characteristics and wants to discover which customer segments exist. There is no known right answer supplied for the groups. Do not reuse the defection answer automatically; the question has changed. Read the options, decide which approach matches this request, and take a moment before revealing your choice. Look for the key difference: this is not asking which customer will receive a known outcome label. It is asking what groups may emerge from the analysis. The method choice follows from that distinction. After you decide, review why each option does or does not fit. Remember, this check selects an approach for the stated goal; it does not establish that the available characteristics are sufficient or that a resulting segment will be useful.

7. Keep the choice narrow: labels guide the first method decision. · 56s

Carry one question into your next analysis conversation: do we already know the outcome labels, or are we trying to discover groups? For a known discrete outcome, such as the left-or-did-not-leave result in our illustrative case, supervised classification fits the request. For customer segments with no supplied right answer, clustering fits the discovery goal. Say the requested result and label status aloud before you name the method; that makes your reasoning visible and easier to check. This is only a first choice among approaches described in the notes. It does not compare every machine-learning method, establish data quality, or guarantee a useful model. This week, use the check on one proposed workplace analysis. Write down its target, whether labels exist, and your choice. Then ask a colleague to challenge whether the chosen approach matches the question.

Sources

  • S1 · Test Admin — coach's own words from an approved lesson.

Attribution: Test Admin. Version 0.1.0.