Documentation
Choosing an ML Task

Choosing a Machine-Learning Task

Start with the output a user needs. “Use AI” is not a task definition; “route this message to one of 12 teams” is.

Translate the outcome into a task

You needChooseExample output
A category from a spreadsheet rowTabular classificationhigh_risk
A number from a spreadsheet rowTabular regression42.7 MPa
A category from textText classificationcard_arrival
Similar wording or likely duplicatesLexical & fuzzy searchRanked matching rows
Metrics over saved responsesResponse evaluationCompletion and token F1
One category for a whole imageImage classificationhealthy_leaf
Objects and their locationsObject detectionLabelled bounding boxes
A class for every pixel or object maskSegmentationPixel masks
Named landmarksKeypoint detectionJoint coordinates

Define the prediction moment

Write one sentence before configuring a model:

Using information available at this moment, predict this output for this kind of case, so that this user can take this action.

This sentence exposes leakage. If a churn model includes cancellation_date, or a loan model includes a field created after approval, the held-out score can look excellent while the deployed model is useless. Mark identifiers and post-outcome fields as excluded features in Tabular AI.

Prefer the simplest useful baseline

A majority-class prediction, average-value regression, lexical search or small linear text classifier gives you a reference point. A more complex model earns its cost only when it improves the metric that matters while meeting latency, privacy and review requirements.

Know when the task needs redesign

  • If labels are subjective, write annotation guidance and measure reviewer agreement.
  • If rare cases matter, choose balanced accuracy, macro F1, recall or another metric that exposes them.
  • If future data differs from historical data, use a time-aware test set outside the current random-split workflow.
  • If the desired output is free-form generation, AnyLearning's current Text AI workflows can organize or evaluate saved text, but they are not an LLM trainer.

Next: evaluate models without fooling yourself.