Prediction, Description, Automation, and Decision Support
Unit ID: ML-M01-U01 Estimated active time: 20-30 minutes
Four different kinds of work
Many requests are called machine learning, but not all of them need a predictive model.
A useful first split is:
| Task type | Main question | Example |
|---|---|---|
| Prediction | What is likely to happen or what label fits this case? | Which learners may need support next week? |
| Description | What is happening in the data? | Which modules have the highest drop-off? |
| Automation | Can a repeated action be handled by a system? | Send a standard reminder after three missed activities. |
| Decision support | Can a person use evidence to make a better decision? | Show mentors a support-priority list with reasons and limits. |
These types can overlap. A prediction may support a decision. A rule may automate part of a workflow. But naming the main type helps us avoid building the wrong thing.
Prediction
Prediction asks for an unknown value.
Examples:
- Will this learner complete the first module?
- What score is this house likely to receive?
- Is this message likely to be spam?
- How many support tickets may arrive tomorrow?
Prediction can be a good fit for supervised machine learning when we have examples from the past with known outcomes.
The word supervised means the training examples include the answer we want the model to learn from. That answer is called the label or target.
Description
Description asks what is already happening.
Examples:
- Which course pages get the most visits?
- Where do learners stop most often?
- What are the common reasons support requests are opened?
- How long do learners usually spend in each module?
Description may need charts, summaries, grouping, or careful data cleaning. It may not need machine learning at all.
This is not a lower-quality result. Sometimes a clear table is more useful than a model.
Automation
Automation asks whether a repeated action can be done by software.
Examples:
- Send a reminder when a learner has not opened the course for seven days.
- Mark a lesson as complete when all required activities are done.
- Route a form submission to the right inbox.
Many automation tasks are best handled by rules. A rule is easier to inspect, easier to explain, and easier to stop.
If the rule works well, do not add machine learning just to make the project sound advanced.
Decision support
Decision support helps a person make a decision.
Examples:
- Give mentors a weekly list of learners who may need help.
- Show a confidence band around a forecast.
- Highlight records that need human review.
Decision support is often where ML is useful. But it also needs care.
The model should not silently replace human judgement when the consequences are serious. It should show enough context for a person to review the result.
Guided practice
Classify each request.
| Request | Best first classification |
|---|---|
| Find which course modules have the most incomplete activities. | Description |
| Predict which learners may need support by the end of week 1. | Prediction |
| Automatically email a link when someone submits the contact form. | Automation |
| Give a mentor a sorted list of learners to review, with reasons and limits. | Decision support |
Now add your own reason for each answer. Do not only copy the label.
The useful question
When you hear a request, ask:
Are we trying to predict, describe, automate, or support a decision?
If the answer is description or simple automation, pause before proposing machine learning.
Takeaway
Machine learning is most natural when we need prediction from examples. But many valuable projects are descriptive reports, rules, or decision-support workflows. In the next unit, we will make a prediction request precise.