What to know before you read.
- Define how a prediction will support a decision before selecting a model.
- Protect evaluation from leakage and compare against a simple baseline.
- Error analysis and limitations make a portfolio project credible.
Write the prediction decision
State who will use the prediction, what action it informs, the target, prediction time and consequence of each error. Check whether the target can be observed without using future information. If no decision changes, a predictive model may not be necessary.
Create trustworthy splits
Profile duplicates, time ordering, groups and data collection. Keep records from the same entity or future period from leaking across splits. Reserve the test set for final evaluation and use validation results for choices.
Establish a baseline
Compare against a simple rule, historical average or lightweight model before tuning. Choose a metric based on error costs and report class or segment-level results. Complexity must earn its place by improving the decision, not just one aggregate score.
Analyse failures
Group false positives and false negatives, review representative records and identify missing features or label problems. Test performance across time or relevant user groups. Explain whether an error requires more data, a different threshold or a human review step.
Make the project reproducible
Provide a data statement, environment, ordered pipeline, experiment record and README. Include model limitations and monitoring signals for a hypothetical deployment. The Machine Learning Certificate of Specialisation connects these decisions in a practical capstone.
Questions readers often ask.
Which dataset should I use?+
Choose a permitted dataset with a clear target, enough context to discuss quality and a manageable size.
How many models should I compare?+
A baseline and two meaningfully different approaches are often enough for a beginner project.
What should a machine-learning README include?+
Include the decision, data, splits, baseline, metric, results, errors, limitations and reproduction steps.
Comparison