Machine Learning for Professionals Training Course

Participants can assess whether a dataset is suitable for a predictive modelling task.

• Analytical datasets structured as rows, features, targets and identifiers
• Data dictionaries document source systems, definitions, refreshes and ownership
• Profiling distributions, missingness, duplicates and invalid category values
• Target labels assessed for timing, completeness and operational consistency
• Sampling frames examined for coverage gaps and selection bias
• Train, validation and test populations defined before modelling begins

Participants can organise a reproducible Python notebook for a machine learning analysis.

• Jupyter Notebook sections separate assumptions, preparation, modelling and findings
• Pandas DataFrames filter, join, group and reshape structured datasets
• NumPy arrays support numerical transformations and model input matrices
• Functions package repeated preparation steps into readable reusable code
• Random seeds make sampling and model results reproducible
• Versioned inputs and outputs create an auditable analysis record

Participants can produce an evidence-based baseline assessment before training complex models.

• Univariate distributions reveal skewness, outliers and impossible numerical values
• Cross-tabulations compare target rates across operationally meaningful segments
• Scatterplots and boxplots expose relationships, ranges and data anomalies
• Correlation analysis identifies redundant variables and potential proxy features
• Majority-class and mean-prediction baselines establish minimum performance thresholds
• Data findings translated into preparation decisions and modelling hypotheses

Participants can prepare model-ready data through repeatable preprocessing pipelines.

• Missing-value strategies selected using data mechanisms and operational meaning
• One-hot encoding converts categorical fields into model-ready feature matrices
• Standardisation and normalisation applied where algorithm assumptions require scaling
• Date fields transformed into recency, seasonality and elapsed-time features
• Outlier treatment justified through domain rules and distribution evidence
• Scikit-learn pipelines prevent inconsistent preprocessing between dataset splits

Participants can train and interpret baseline classification models for binary outcomes.

• Logistic regression estimates event probabilities from structured predictor variables
• Decision trees create interpretable rule-based classification models
• k-nearest neighbours classify observations through local similarity patterns
• Class imbalance quantified using prevalence, counts and minority-class representation
• Confusion matrices distinguish true positives, false positives and missed cases
• Decision thresholds aligned with intervention capacity and error costs

Participants can build regression models and assess forecast error for continuous targets.

• Linear regression models continuous outcomes using additive predictor relationships
• Regularisation with Ridge and Lasso controls unstable coefficient estimates
• Residual plots test systematic error across prediction ranges
• MAE, RMSE and MAPE compared for operational forecasting decisions
• Prediction intervals communicate uncertainty around individual estimated values
• Forecast errors segmented by region, product and customer group

Participants can create and select defensible features without contaminating model evaluation.

• Domain features derived from transactions, interactions and historical event sequences
• Aggregation windows calculate counts, averages and recent behavioural changes
• Interaction terms represent combined effects between relevant predictor variables
• Leakage checks identify information unavailable at prediction time
• Filter methods rank features using variance, correlation and mutual information
• Feature documentation records formulae, sources, timing and business rationale

Participants can choose validation designs that reflect intended model use.

• Holdout validation separates development evidence from final performance evidence
• k-fold cross-validation estimates performance variation across multiple data partitions
• Stratified splits preserve event prevalence in classification datasets
• Time-based splits prevent future observations entering historical prediction models
• Grouped validation prevents related records appearing across training and test sets
• Model comparison tables report metrics, variability, assumptions and trade-offs

Participants can compare ensemble models with simpler baselines using justified performance criteria.

• Random forests combine diverse decision trees to reduce prediction variance
• Gradient boosting sequentially corrects errors from earlier weak learners
• XGBoost parameters control tree depth, learning rate and regularisation
• Feature importance scores interpreted cautiously alongside correlation and proxies
• Overfitting diagnosed through training-validation performance gaps and learning curves
• Complexity choices balanced against interpretability, latency and maintenance costs

Participants can tune model pipelines and record experiments for repeatable comparison.

• Grid search evaluates defined parameter combinations across validation folds
• Random search samples broader parameter spaces under limited computing budgets
• Bayesian optimisation prioritises promising configurations from earlier trial results
• Nested cross-validation separates tuning decisions from final performance estimates
• Experiment logs capture data versions, parameters, metrics and decisions
• MLflow records model runs, artefacts and selected model versions

Participants can convert model scores into decisions using calibrated probabilities and costs.

• Predicted probabilities distinguished from ranked risk scores and class labels
• Calibration curves compare estimated probabilities with observed event frequencies
• Platt scaling and isotonic regression calibrate classification probability outputs
• ROC curves assess discrimination across alternative classification thresholds
• Precision-recall curves focus evaluation on rare-event prediction performance
• Cost matrices select thresholds using intervention economics and capacity limits

Participants can identify segments and unusual cases using appropriate unsupervised methods.

• k-means clustering groups observations around similarity-based centroid locations
• Hierarchical clustering reveals nested group structures through linkage distances
• Silhouette scores assess cluster separation and within-cluster cohesion
• Principal component analysis reduces correlated features for visual inspection
• Isolation forests identify unusual records through random partitioning patterns
• Cluster profiles translated into service, targeting and investigation actions

Participants can explain model behaviour to technical and non-technical stakeholders.

• Global feature importance summarises variables associated with model predictions
• Partial dependence plots show average prediction changes across feature values
• SHAP values explain feature contributions for individual model predictions
• Local explanations tested against data quality and plausible causal mechanisms
• Surrogate models provide simplified approximations of complex prediction functions
• Interpretation statements distinguish association, prediction and causal claims

Participants can identify key governance risks and propose proportionate model controls.

• Protected characteristics identified within legal, ethical and policy constraints
• Demographic parity compares positive prediction rates between defined groups
• Equal opportunity compares true positive rates across relevant groups
• Proxy discrimination assessed through correlated features and outcome disparities
• Privacy risks reviewed through minimisation, access controls and retention periods
• Model risk registers assign owners, controls, triggers and review dates

Participants can specify an operational deployment and monitoring approach for a machine learning model.

• Batch scoring and real-time APIs compared for operational use cases
• Inference pipelines reproduce approved preprocessing and feature transformations
• Model cards document intended use, limits, metrics and governance decisions
• Data drift monitored through feature distributions and population stability indices
• Performance drift detected through delayed outcomes and periodic back-testing
• Human review routes define overrides, exceptions and feedback capture

Participants can build, review and present a complete machine learning project recommendation with reproducible evidence.

• Business problem framed with target, decision, users and success measures
• Dataset assessed for quality, coverage, labels and potential prediction risks
• Reproducible pipeline prepared with documented features and validation design
• Candidate models compared using calibrated metrics and business cost evidence
• Fairness, explainability and deployment controls reviewed before recommendation approval
• Project findings presented through model report, implementation plan and stakeholder briefing