Applied Machine Learning

Translate a maintenance problem into a target, prediction time and action.

Evidence: A brief excludes information unavailable at prediction time.

Inspect class balance, distributions and missing values.

Evidence: An EDA notebook identifies at least three relevant data risks.

Create stratified or time-aware splits appropriate to the scenario.

Evidence: A leakage audit explains split choice and holds test data untouched.

Fit imputers, encoders and scalers on training data only.

Evidence: A pipeline transforms validation data without refitting.

Implement a simple baseline and a linear or logistic model.

Evidence: Results include a baseline and justified task-specific metrics.

Compare a decision tree and a random forest.

Evidence: A controlled comparison uses the same split and records parameters.

Compare class weights and training-only resampling.

Evidence: Validation distribution remains unchanged and minority metrics are reported.

Use cross-validation and bounded hyperparameter search.

Evidence: Search decisions do not use the held-out test set.

Choose an operating threshold using costs and validation evidence.

Evidence: Precision, recall and expected workload are reported at that threshold.

Inspect feature influence and subgroup error patterns.

Evidence: A limitation report avoids causal claims from feature importance.

Save preprocessing and model together behind a local API.

Evidence: Predictions match notebook results on fixed test inputs.

Define data drift, performance review and rollback triggers.

Evidence: A model card and runbook identify owners and data limitations.