Why a Model Is Not the First Question

Unit ID: ML-M01-U00 Estimated active time: 15-20 minutes

Let us begin

Imagine that someone comes to you and says:

We have course data. Can you build a machine-learning model?

That sounds like a clear request, but it is not clear yet.

We do not know what the model should predict. We do not know who will use the result. We do not know what action will be taken. We do not know what mistake would be serious.

So the first answer should not be:

Yes, let us try logistic regression.

The first answer should be:

What decision are we trying to improve?

This is the starting habit for the whole course.

Algorithm-first thinking is risky

Machine learning can feel exciting because there are many algorithms. But an algorithm is only useful when the problem is framed well.

If we start with the algorithm, several things can go wrong.

• We may predict something that no one can act on.
• We may use data that will not be available at prediction time.
• We may create a score that looks accurate but does not help the real decision.
• We may harm people by using a weak or unfair signal.
• We may spend time building a model when a simple rule or report is enough.

Good ML work begins before code.

A classroom example

For this module, we will use a fictional example:

An online course team wants to support learners who may not finish the foundation course.

This is still too broad. It sounds reasonable, but we need to ask more.

Do we want to predict final completion? Do we want to find learners who need help this week? Do we want to describe why learners leave? Do we want to automate reminder emails?

Each version is different. Each version needs different data, different evaluation, and different safeguards.

The central question

Before choosing a model, ask:

What useful action will become better if this prediction is available?

If there is no clear action, the model may become a number that people admire but do not use.

For the learner-support example, a useful action could be:

A mentor checks in with a learner during the first week if the learner appears likely to need support.

Now the project is becoming clearer. We are not just predicting for fun. We are trying to improve a support action.

Prediction is not explanation

A machine-learning model can often predict that something is more likely.

It does not automatically explain why it happened.

For example, a model may learn that learners with fewer practice attempts are less likely to finish. That does not prove that low practice attempts caused non-completion. Maybe those learners had less time, worse internet, unclear instructions, or another problem not recorded in the data.

In this course, we will be careful with language:

• The model predicts a higher risk is a prediction statement.
• This factor caused the learner to stop is a causal statement.

Do not turn the first statement into the second without stronger evidence.

Quick check

Read each request and decide whether it is ready for modelling.

1. Build a model that improves learner support.
2. Predict by the end of day 3 whether a learner is likely to complete Module 1, so a mentor can offer help on day 4.
3. Use AI to understand learners better.

The second request is the closest to ready. It names a prediction time, an outcome, and an action. It still needs more detail, but it is much stronger than the other two.

Takeaway

A model is not the first question. The first question is the decision or action we want to improve. In the next unit, we will separate four common task types: prediction, description, automation, and decision support.