Problem Definition
How to convert a business objective into a well-posed data science problem: the target variable, unit of analysis, success criteria, and a preliminary plan.
3 min read · Updated August 8, 2026
Problem definition means understanding the project objectives and requirements from a domain perspective, then converting that knowledge into a data science problem definition with a preliminary plan designed to achieve the objectives. It is the first stage of the Data Mining Map, and the one that most determines whether the rest of the project succeeds.
A successful data science project starts from a well-defined question or need. “Use our data to improve the business” is not a question; it is a wish. The work of this stage is turning wishes into specifications.
From business objective to data problem
A well-posed problem definition answers five questions:
- Objective — what business outcome are we trying to change? (Reduce churn, catch fraud, forecast demand.)
- Target — what quantity, exactly, will we describe or predict? This becomes the target variable of modeling.
- Unit of analysis — what is one observation? A customer, a transaction, a patient visit, a day?
- Scope and data — what populations, time windows, and data sources are in bounds?
- Success criteria — how will we know the result is useful? A business metric (dollars saved) and a model metric (see model evaluation) are both required.
The output is a preliminary plan: the exploration, preparation, and modeling steps you expect to need, the people who own them, and the decision the finished model will inform.
Where data science gets applied
The legacy site illustrated the breadth of problem domains with a KDnuggets reader poll on where analytics, data mining, and data science were applied. The leading sectors, preserved from that poll:
| Industry | Share of voters |
|---|---|
| CRM / consumer analytics | 16.3% |
| Finance | 15.0% |
| Banking | 13.4% |
| Advertising | 12.0% |
| Science | 12.0% |
| Health care | 12.0% |
| Fraud detection | 11.1% |
| Retail | 10.3% |
| Insurance | 9.2% |
| E-commerce | 8.9% |
The long tail — telecom, social media, credit scoring, biotech and genomics, manufacturing, education, energy, government — matters as much as the leaders: the process you are learning here is the same in every one of them. Only the domain vocabulary changes.
In practice
Modern teams write the problem definition down as a one-page brief or design doc before touching data, and revisit it at every iteration of the CRISP-DM loop. Framing choices made here — classification vs. regression, what counts as a positive — directly determine which scikit-learn estimator, which evaluation metric, and which threshold you will use later. A common modern addition is an explicit fairness and privacy review at this stage, before data collection.
Common pitfalls
- Starting with a technique (“let’s use deep learning”) instead of a question.
- A target variable that doesn’t match the business objective — predicting clicks when the goal is revenue.
- No agreed success criteria, so “good enough” is litigated after the model is built.
- Defining the problem around the data you have rather than the decision you need to make.
- Skipping the preliminary plan, so exploration and preparation have no time budget.
Summary
Problem definition converts a domain objective into a data science specification: objective, target, unit of analysis, scope, and success criteria, plus a preliminary plan. Get this stage right and every downstream stage — exploration, preparation, modeling, evaluation, deployment — has a target to aim at.