Features, Labels and Models
Features describe each example; labels state what a supervised model should predict; the model learns a relationship between them.
Learning objectives
What you will be able to do
- Explain features, labels and models accurately in your own words.
- Recognise how the concept appears in real AI products and professional work.
- Identify an important limitation, risk or evaluation requirement.
- Apply the concept through a practical activity and knowledge check.
Core explanation
Build the right mental model
Feature engineering converts raw information into signals a model can use. A date might become day of week, a network log might become connection rate, and free text might become an embedding.
Features should be available at prediction time, reliably measured and appropriate for the decision. A proxy feature can reproduce sensitive attributes and create indirect discrimination.
A model is the learned function, not the complete AI system. The system also includes data collection, preprocessing, thresholds, interfaces, monitoring and human procedures.
Write a prediction contract
A feature is an input representation supplied to a model. A label is the target attached to a supervised training example. A prediction is the output produced for a case. The model is the mapping that connects inputs to outputs. Keeping these terms separate prevents a common mistake: including the answer, or information derived from the answer, among the inputs.
A prediction contract states the unit of observation, prediction time, available inputs and target definition. For example: “For one support request at the moment it is submitted, estimate hours until first staff response.” The unit is a request, the time is submission and the label is a later duration. That contract immediately tells you that the eventual response timestamp cannot be an input.
Turn raw records into defensible features
Raw records often need representation. A date can become a day-of-week category; a text description can become a numerical representation; a missing reading can need an explicit missing-value treatment. Each transformation makes assumptions. A numeric code for a category should not silently imply that category 3 is three times category 1.
Use a data dictionary containing the feature’s meaning, units, source, valid range, missing-value rule and availability time. This helps reveal mismatched units, unavailable inputs and ambiguous codes. A feature that is convenient in a retrospective spreadsheet may be impossible to obtain consistently when the live prediction is required.
Check suitability as well as predictive usefulness
A feature can correlate with a target while being inappropriate for the intended use. It may expose unnecessary personal information, reflect unequal access to services or encode a proxy for a sensitive characteristic. Removing a direct identifier does not guarantee that the remaining inputs cannot identify a person or reproduce an unwanted distinction.
Also separate the model from the application around it. The model may estimate a response time, while the application decides what to show the requester. The surrounding system must handle missing inputs, unavailable services, monitoring and corrections. A model file alone cannot specify all of those operational behaviours.
Example
One record or case supplied to a model.
Feature
An input attribute used to make a prediction.
Label
The target outcome in supervised learning.
Model
The learned mapping from features to an output.
Build a feature table for a fictional service desk
At submission time, suppose a request has category “network,” description length 80 words and queue length 12. Its eventual first response arrives three hours later. The category, text length and current queue length are candidate features; the three-hour duration is the training label. The responder’s later identity and resolution summary are not available at submission and should not be included for this prediction.
Consider an invented model: predicted hours = 0.5 + 0.1 × queue length. For a queue length of 12, it predicts 1.7 hours. Against the three-hour label, the error is 1.7 − 3 = −1.3 hours: an underestimate. The model does not use category or description length even though the record contains them; a stored field becomes a model feature only when it is actually supplied to the model.
If queue length is missing, substituting zero would imply an empty queue, which is a different statement. The application needs a documented missing-input policy, such as withholding an estimate or using a separately evaluated fallback. Do not invent certainty merely to fill the interface.
Apply your learning
Practical activity
- Spend five minutes writing a prediction contract for a fictional task in your profession.
- Spend five minutes listing six candidate fields with units and availability times.
- Spend five minutes separating usable features, labels, identifiers and future information.
- Spend five minutes calculating the toy service-desk model’s estimate for queue length 5 and writing a missing-input response.
Compare with an explained example answer
For queue length 5, predicted hours = 0.5 + 0.1 × 5 = 1.0 hour. This follows the invented formula; it does not prove that real requests will receive a response in one hour.
A suitable missing-input response is “An estimate is unavailable because the current queue count is missing.” A task-specific data dictionary should explain which fields are available before the decision and why each is needed, rather than simply including every column in a spreadsheet.
Write your answers in a learning journal. The value comes from connecting the concept to your own profession.
Knowledge check
Test your understanding
Q1Is the eventual response time a feature for a prediction made at submission?
Answer: No. It becomes the label once observed; using it as an input would leak future information.
Q2Is every stored field a model feature?
Answer: No. Only fields or derived representations actually supplied to the model are features.
Q3Why is replacing a missing queue count with zero risky?
Answer: Missing means unknown; zero means the queue is empty. Treating them as equivalent can create a misleading prediction.
If you cannot explain the answer without reading it, revisit the core explanation and example before continuing.
Continue with authoritative sources
Important industry resources
These links lead to official organisations, industry laboratories or established open-source learning projects.