Does data-driven mean deep learning?

No. Linear regression, trees and many other methods can be data-driven.

Can it use expert knowledge?

Yes. Experts can shape inputs, constraints and interpretation.

Does it explain why an outcome happens?

Not necessarily. Predictive success alone does not provide a causal explanation.

AI Glossary / AI and machine learning / Beginner

Data-driven model

Definition

A data-driven model is a model whose relationships or parameters are learned substantially from observed data. It uses examples to estimate patterns, often for prediction, classification or control, rather than relying only on equations specified from first principles.

What it means

A demand model can learn how past sales relate to season, price and promotions. Its behavior reflects the examples and modeling choices supplied to it. It does not automatically discover the underlying causes of demand.

Data-driven and knowledge-driven approaches are not absolute opposites. A model can combine physical equations, expert constraints and learned components. The useful question is which parts come from data and which assumptions are imposed.

For a business, the label should prompt questions about data relevance. A large historical dataset may be unsuitable if it describes a different product, population or process from the one the model will face.

How it works

The team defines a target or task, assembles suitable observations and chooses a model family. Training estimates parameters or structure by optimizing a learning objective.

Validation guides development, while a separate evaluation assesses performance on relevant unseen cases. The data split should reflect the intended use, particularly when time, repeated entities or changing conditions matter.

After deployment, input quality and outcomes are monitored. A learned relationship can weaken when behavior, measurement or the environment changes. Updating the model requires a controlled process rather than an assumption that more recent data is automatically better.

Three practical examples

A retailer predicts demand from historical sales and contextual variables, while accounting for stockouts that hide unmet demand.
A service organization classifies requests using previously labeled examples.
An engineering team learns a residual correction to a physical model, combining measured data with known system structure.

Think of it this way

Think of learning the journey time to work from many previous trips. The observations help forecast the next trip, but a road closure can invalidate the old pattern.

The analogy has limits because a person may know why conditions changed. A model only reacts to the information and update process provided to it.

Worked example

Prediction is not the same as intervention

Suppose a fictional retailer observes that high sales often coincide with large advertising budgets. A model trained on those records may predict higher sales when the budget field is larger.

That relationship does not establish how much sales would increase if the retailer changed the budget. Managers may spend more when they already expect strong seasonal demand. The data contains both the advertising decision and the circumstances behind it.

The model can still be useful for a clearly defined forecasting task if evaluated properly. For a budget-allocation decision, however, the team needs evidence suited to estimating the effect of intervention, such as a well-designed experiment or an appropriate causal analysis.

The example distinguishes two questions that can look similar in a spreadsheet: what tends to occur together, and what would change if the business acted differently.

Illustrative case study

A fictional repair business trains a model to estimate job duration. Historical records show short times for certain jobs because technicians previously excluded waiting-for-parts time from that field.

The business now wants end-to-end completion estimates for customers. The old label no longer matches the new question. The team revises its measurement definition and gathers suitable outcomes before evaluating a replacement model.

The case illustrates why data-driven does not mean assumption-free. The model learns the recorded target, which may differ from the operational outcome stakeholders think they are receiving.

Common misconceptions and mistakes

Assuming a model trained on data is objective by default: collection and labeling choices influence it.
Treating predictive relationships as causal effects: intervention questions need appropriate evidence.
Using more rows without checking relevance: quantity cannot repair a mismatched target or population.
Believing first-principles knowledge is incompatible with learning: hybrid models can use both.

Business use

Data-driven models are useful when historical examples contain repeatable information relevant to a task. Their value depends on whether that information remains available and reliable when decisions are made.

Ask what the target means, which cases the dataset excludes and whether each input is available at prediction time. Compare the model with a simple baseline and test changed conditions where feasible. Keep a clear owner for monitoring and updates, because deployment is the start of an operating process rather than the end of modeling.

Check your understanding

A model predicts repair completion using a field recorded only after the repair is finished. Is that a useful input for a prediction made when the job arrives?

Answer: No. It is unavailable at the intended decision time and can create misleading evaluation results through leakage.

Related terms

Learn next

Read generalization. Write the target, prediction time and available inputs for one proposed model, then check whether the historical records match that specification.

Generalization

Sources

NIST/SEMATECH, Exploratory Data Analysis: https://www.itl.nist.gov/div898/handbook/eda/section1/eda11.htm, contextual overview consulted. Scikit-learn, Common pitfalls: https://scikit-learn.org/stable/common_pitfalls.html, sections 12.1-12.2 checked for consistent preprocessing, prediction-time availability and separation of training and test data. This broad entry uses original explanations rather than claiming one universal formal definition.

Previous: Data drift Next: Data editing