Practical Applications of Machine Learning
Practical Applications tests whether you can take a model out of a notebook and make it useful and safe in an organisation. It covers problem framing, data governance, deployment patterns, monitoring, and ethical and legal constraints.
From problem to system
- Frame the business objective as a measurable ML task. 'Reduce customer churn' becomes 'predict probability of cancellation in the next 60 days', with a decision rule and a baseline (current rule or majority class) to beat.
- Choose the metric from the cost of errors, not convention. In loan default, a false negative (bad loan approved) may cost far more than a false positive (good applicant reviewed).
- Data governance: collect only necessary data, obtain consent, minimise and anonymise personal data, and comply with GDPR rights such as access and explanation. Document provenance and quality.
- Deployment patterns: batch scoring (nightly churn list), real-time API (fraud check at payment), or embedded on-device. Version the model, features, and training data.
- Monitoring: watch input drift (feature distributions change), concept drift (the relationship between features and target changes), and live performance where labels arrive late. Define retraining triggers and a rollback path.
- Ethics and explainability: test for bias across groups, provide human review for high-impact decisions, and use tools such as SHAP or LIME to explain individual predictions. Accountability stays with the organisation.
Worked example
A bank deploys a loan-default model. Offline, it reaches recall 0.75 at a 70% approval rate, against a 2.5% baseline default rate. Before launch, the team checks approval rates across age bands and finds a 15-point gap, so they remove a proxy feature and revalidate. They deploy as a real-time API, shadow-score for a month, then canary-release to 5% of applicants with human review of rejections. Monthly monitoring tracks drift and recall; retraining is triggered if recall drops below 0.70 for two consecutive months.
Traps
Optimising offline accuracy without linking it to business cost. Assuming offline performance survives deployment. Launching without monitoring or a rollback plan. Ignoring fairness and privacy because the model is 'just statistics'. Skipping a baseline, so improvement cannot be proven.