Machine learning: the tasks it can address and the data it needs

Machine learning finds patterns to predict values or classify new cases. It needs a clear task and suitable history. Many records do not automatically form a useful dataset; correct targets, quality and relevance to future use matter.

Define the prediction and decision time

Agree what is predicted and when the process needs it. Demand, risk and request classification need different data and criteria. Compare models with a baseline rule to justify complexity.

For example, demand prediction must use information available before purchasing. Including post-sale data can make tests impressive but unrealistic. Feature availability timing belongs in requirements.

Validate data and target labels

Assess missing values, duplicates, definition changes and target generation. For classification, consistent human labelling matters. Unstable labels teach contradictions, so define rules and review examples.

Agree access and permitted use of sensitive data. Do not add fields merely because they exist. Justify usefulness, quality and future availability for the task.

Evaluate independent examples

Separate training and evaluation according to task meaning, avoiding leakage of the same cases. Temporal processes require event ordering. Repeatedly tuning against the final test can make conclusions overly optimistic.

Metrics depend on error costs and class distribution. Overall accuracy may hide failure to detect a rare important class. Review missed cases and false positives separately with the process owner.

Monitoring is needed after the pilot

Pilots assess integration, latency and human action based on predictions. Define behaviour when data is missing or results are unsuitable. Models should belong to a managed process rather than an opaque decision without investigation.

Assortment, customer behaviour or source changes can reduce quality. Retain versions, control examples and review procedures. Retraining also needs evaluation; new data do not automatically guarantee a better model.

Sources and further reading

• Google: dataset splitting
• Google: classification metrics
• NIST: AI risk management

• Related subservice: Machine learning
• Explore this service area