What it is
Supervised learning fits a model to examples that pair input data with known labels or outcomes. It is commonly used for classification and regression. City teams can use it to estimate travel time, classify service requests, or predict building loads, provided the training examples are representative and the output supports a defined decision.
How it works
- 01Define the prediction target and decision that will use it.
- 02Assemble lawful, relevant examples and document how labels were produced.
- 03Separate training data from later or geographically held-out evaluation data.
- 04Compare performance with a simple baseline and inspect errors across neighborhoods, seasons, and user groups.
- 01
Define the prediction target and decision that will use it.
- 02
Assemble lawful, relevant examples and document how labels were produced.
- 03
Separate training data from later or geographically held-out evaluation data.
- 04
Compare performance with a simple baseline and inspect errors across neighborhoods, seasons, and user groups.
Why it matters
Labels encode institutional choices as well as observations. A model trained on past enforcement or service records may reproduce where attention was previously concentrated. Evaluation should test the data-generating process, not only the model score.
Typical Smart City applications
- Classifying transport or maintenance events
- Predicting energy and water demand
- Estimating asset condition from inspections
- Routing service requests for staff review
Example data inputs
Strengths
- Supports clear prediction tasks with measurable outcomes
- Can reuse historical examples when definitions are stable
- Enables error analysis against known outcomes
Limitations and risks
- Biased or incomplete labels can encode historic service inequities.
- Distribution changes can make old examples poor guides to new conditions.
- A predictive association does not establish a cause or justify an intervention.
Illustrative workflow
Applied city example
A city trains a model to triage pothole reports using photos and inspection outcomes. It checks error rates by lighting, road type, and neighborhood, keeps staff responsible for dispatch, and reviews whether the training labels reflect the city's current service standard.
Related methods
- Classification algorithms
- Regression algorithms
- Anomaly detection
Frequently asked questions
Do labeled records need to be perfect?
They need documented definitions and enough quality for the intended task. Label disagreement and uncertainty should be measured.
Can a model trained in one city be transferred to another?
Not without checking differences in sensors, policy, population, geography, and operating practice.
Does a good test score mean a useful city service?
No. The score must connect to a real workflow and be balanced against error costs, equity, and maintenance burden.