Problem definition translates a real-world objective into a precise machine learning problem with suitable inputs, outputs, evaluation criteria, and operational constraints. It starts the ML life cycle and must align stakeholders before data collection or model development begins. section 2.1 section 2.0
A production-ready definition answers questions across three stakeholder groups:
- Business and product teams clarify the task, reason for using ML, success metrics, business value, timeline, error tolerance, and impact of incorrect predictions. section 2.1
- Technical teams assess whether ML is appropriate, identify available data and pipeline needs, choose evaluation metrics, estimate compute, and plan deployment and retraining. section 2.1
- Legal and compliance teams address ethics, privacy, sensitive information, and data-governance requirements. section 2.1
From vague goal to defined problem
“Predict the churn rate of customers” is insufficient by itself because it does not establish the task details or evaluation criteria. A vehicle-price system instead connects prediction to customer needs, business value, evolving data, and future model versions. section 2.1
Figure 2.2 ML as a loop
Use the ML-as-a-loop figure to place problem definition at the start of an iterative workflow that continues through data, modeling, deployment, and improvement. section 2.1
Use the mental map to connect defining the problem with model design and the input requirements that constrain later development. section 2.0
