Feasibility: Data, Labels, Timing, and Cost

Unit ID: ML-M01-U05 Estimated active time: 25-35 minutes

A good idea may still be impossible

A problem can be important and still not be ready for machine learning.

Before modelling, check feasibility.

Feasibility means:

Do we have enough suitable evidence, at the right time, to build and use the model responsibly?

This is not only a technical question. It is also a practical and ethical question.

Labels

For supervised learning, we need past examples with known answers.

In the learner-support example, a possible label is:

completed_module_1_by_day_10

Ask:

• Is this label recorded?
• Is it recorded consistently?
• Does it mean what we think it means?
• Is it available for enough past learners?
• Could the label itself be unfair or misleading?

If the label is weak, the model learns from weak evidence.

Prediction-time data

A model can only use data available at prediction time.

If we predict at the end of day 3, the model can use:

• day 1 to day 3 activity;
• number of attempts so far;
• whether the learner opened required lessons;
• support requests made before day 3.

It cannot use:

• day 10 completion;
• later quiz scores;
• future support messages;
• final course feedback.

Using future information is one type of leakage. Leakage makes a model look better during testing than it can be in real use.

Data volume and coverage

Ask:

• How many past examples are available?
• Do the examples cover different learner situations?
• Are some groups missing or under-recorded?
• Has the course changed since the data was collected?

More data is not automatically better. Badly measured data at large scale can still teach the wrong pattern.

For this first course, we will use small fictional or synthetic data so the workflow is clear. In real work, this check becomes much more serious.

Cost and action capacity

A prediction is useful only if the action is possible.

Ask:

• Who will act on the model result?
• How much time does the action take?
• What happens if many cases are flagged?
• Is there a safe fallback when the model is uncertain?
• Who is responsible for stopping the system if it behaves badly?

For the learner-support example, a mentor may be able to send 20 careful messages per week, but not 500. That capacity affects the model threshold and the whole design.

Feasibility decision

Use three categories.

Stopping is not failure. It is good project judgement.

Guided practice

Review this frame:

At the end of day 3, predict whether a learner is likely to complete Module 1 by day 10, so a mentor can offer optional support on day 4.

Write short answers.

Takeaway

Feasibility depends on labels, prediction-time data, coverage, action capacity, and risk. In the next unit, you will combine everything into a first problem brief.