Target, Unit, Time Horizon, and Action

Unit ID: ML-M01-U02 Estimated active time: 25-35 minutes

A vague goal is not enough

Consider this request:

Predict learner success.

It sounds useful. But a model cannot train on the word success. We must define what success means in data.

Does success mean finishing the whole course? Passing a knowledge check? Completing Module 1? Returning next week? Reporting confidence?

Each meaning creates a different target.

Four parts of a strong prediction frame

A strong prediction frame names four things:

1. Target: what the model predicts.
2. Unit of analysis: one prediction is made for what kind of thing.
3. Time horizon: when the prediction is made and what future period it covers.
4. Action: what someone will do with the prediction.

Let us use the learner-support example.

Weak frame:

Predict learner success.

Stronger frame:

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

Now we can inspect the project.

Target

The target is the answer the model learns to predict.

In our stronger frame, the target is:

Did the learner complete Module 1 by day 10?

This could be recorded as completed_module_1_by_day_10, with values like yes and no, or 1 and 0.

The target must be clear enough that two people would label the same record the same way.

If the target depends on private judgement, vague scoring, or data that is not recorded consistently, the project is weak.

Unit of analysis

The unit of analysis is what one row or one prediction represents.

Examples:

• One learner.
• One learner-module attempt.
• One lesson view.
• One support ticket.
• One purchase.

For our example, one prediction is for:

One learner's first attempt at Module 1.

This matters. If the same learner can appear many times, we must handle that carefully later. If the row is a lesson view instead of a learner, the target and action change.

Time horizon

The time horizon answers two timing questions:

• When do we make the prediction?
• What future outcome are we predicting?

For our example:

• Prediction time: end of day 3 after the learner starts Module 1.
• Outcome window: completion by day 10.

Timing protects us from a common mistake: using future information.

If we predict on day 3, we cannot use data from day 4 or day 10 as an input. That would make the experiment look strong but fail in real use.

Action

The action is what changes because the prediction exists.

For our example:

A mentor sends a short support message on day 4.

This action should be realistic, ethical, and possible.

If mentors cannot actually contact learners, or if the message would be annoying or harmful, the prediction is not enough.

Guided rewrite

Rewrite this vague goal:

Use ML to improve course completion.

A better version:

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

Now check the four parts.

Your turn

Rewrite one of these.

1. Predict which customers are valuable.
2. Predict which learners need help.
3. Predict which support tickets are urgent.

Use this sentence frame:

At [prediction time], predict whether [unit] will [target] by [future time], so [user] can [action].

Takeaway

A useful ML problem frame names the target, unit, time horizon, and action. Without these, we do not yet have a modelling problem. In the next unit, we will ask who uses the result and who may be affected by it.