ML Problem Framing
Master ML problem framing: assess when ML is appropriate, structure solutions, define success metrics, and design effective ML strategies.
What is ML problem framing?
ML problem framing is the process of deciding whether machine learning is the right tool, translating a business need into a learnable task, and defining how success will be measured. It matters because a technically strong model still fails when it solves the wrong problem, relies on unavailable data, or optimizes a metric that does not improve the product.
Before this lesson, review Machine Learning Fundamentals if terms such as model, feature, label, and inference are unfamiliar.
Ask four questions first:
- Is ML appropriate for this problem?
- What kind of ML task represents the desired decision?
- Which business and model metrics define success?
- What data, latency, cost, safety, and operational constraints apply?
When is ML appropriate?
ML is appropriate when a system must learn patterns from examples and can tolerate probabilistic outputs. A deterministic rule-based system is usually better when the rules are simple, stable, and required to be perfectly explainable.
Good fit for ML
- Pattern recognition: the decision depends on complex patterns in data.
- Large datasets: enough representative training examples are available.
- Predictive value: better predictions create measurable business value.
- Automation benefits: the decision must be made at a scale humans cannot handle.
- Tolerance for errors: the workflow can manage false positives and false negatives.
- Changing patterns: learned behavior can adapt more economically than hand-written rules.
Poor fit for ML
- Simple rules: basic logic solves the problem reliably.
- Limited data: there are too few examples or labels.
- High stakes: the workflow has zero tolerance for model mistakes and no human fallback.
- Regulatory constraints: every outcome requires a complete deterministic explanation.
- Static problems: the rules rarely change.
- Small scale: a manual or rules-based approach is sufficient.
Problem Suitability Assessment
Select a scenario to compare its requirements with the conditions that make ML useful.
Excellent fit
Clear patterns, labeled data, high business value, and achievable real-time requirements make this a strong ML problem.
What types of ML problems are common?
An ML problem type defines the shape of the output the model must learn. Choosing the type early determines the labels, metrics, and candidate algorithms.
Classification
Classification predicts one of a set of categories.
- Examples: spam detection, image recognition, and sentiment analysis.
- Success metrics: accuracy, precision, recall, F1 score, and task-specific costs.
- Data requirement: labeled examples for each class.
Regression
Regression predicts a continuous numerical value.
- Examples: house prices, demand forecasts, and delivery time estimates.
- Success metrics: RMSE, MAE, and R-squared.
- Data requirement: historical inputs paired with target values.
Clustering
Clustering groups similar examples without predefined labels.
- Examples: customer segmentation, anomaly exploration, and document grouping.
- Success metrics: silhouette score plus validation that the groups are useful to the business.
- Data requirement: unlabeled examples with features that represent meaningful similarity.
How should success be defined?
Success requires both a business outcome and an ML measurement. The ML metric is a proxy used during development; the business metric verifies that the deployed system creates value.
Confirm the framing is decision-ready
Do not accept a single aggregate metric without examining error costs. For fraud detection, for example, false negatives lose money while false positives inconvenience legitimate customers. The operating threshold must balance both outcomes.
How is a problem framing document structured?
A framing document records the business problem, model task, available data, constraints, baseline, metrics, and deployment assumptions so product, data, and engineering teams can evaluate the same proposal.
What are common framing pitfalls?
Framing pitfalls are assumptions that make a project look feasible before its data and operating conditions have been tested.
Solution-first thinking
Starting with "we need deep learning" instead of identifying the business decision and comparing ML with simpler approaches.
Ignoring data reality
Assuming the right data exists without checking its quality, completeness, labels, permissions, and relevance to production.
Unrealistic expectations
Demanding 99% accuracy when 85% would create substantial value, or when the available labels cannot support a reliable 99% estimate.
Ignoring implementation
Failing to plan how predictions will be served, reviewed, monitored, maintained, and integrated with existing systems.
What comes after framing?
Once the task and success criteria are explicit, continue through the implementation sequence:
- ML Systems Design for architecture and infrastructure planning.
- Advanced ML Data Preparation for collecting and preparing datasets.
- Feature Engineering for creating useful model inputs.
- Model selection and baseline evaluation before investing in a complex production model.