Machine Learning Model Development Training Course

Participants can organise a reproducible Python workflow for machine learning development.

• Python environments, package management and dependency version control
• Jupyter notebooks structured for reproducible analytical and modelling work
• pandas DataFrames for joins, filtering, aggregation and missing-value inspection
• scikit-learn estimator interface, pipelines, transformers and model persistence
• Configuration files separating parameters, paths and environment-specific settings
• Git branching, commits and pull requests for collaborative model development

Participants can assess dataset fitness, identify quality issues and define remediation actions.

• Column profiling for types, ranges, cardinality, completeness and duplicates
• Missingness mechanisms: MCAR, MAR and MNAR assumptions examined
• Outlier detection using z-scores, IQR rules and isolation forests
• Data lineage mapped from source systems through analytical datasets
• Temporal consistency checks for timestamps, reporting periods and late arrivals
• Data quality scorecards with thresholds, owners and remediation actions

Participants can create preprocessing pipelines that transform raw fields into model-ready features.

• Train-test separation before fitting imputers, encoders and scaling transformations
• Numerical imputation using median, KNN and iterative imputation methods
• Categorical encoding with one-hot, ordinal, frequency and target encoding
• StandardScaler, RobustScaler and MinMaxScaler selected by distribution characteristics
• Feature pipelines built with ColumnTransformer and Pipeline objects
• Transformation metadata retained for reproducibility, audit and deployment

Participants can identify predictive patterns, confounders and data limitations before model selection.

• Univariate distributions examined using histograms, density plots and quantiles
• Target rate analysis across customer, product, channel and period segments
• Correlation matrices interpreted alongside nonlinear relationships and multicollinearity
• Scatterplots and boxplots used to identify candidate predictive relationships
• Cohort analysis reveals drift, seasonality and changing population composition
• Data leakage risks identified from post-outcome fields and proxy variables

Participants can engineer meaningful numerical, categorical, time-based and interaction features.

• Domain features translated from business rules, ratios and behavioural signals
• Date features derived from recency, frequency, duration and seasonality
• Interaction terms capture conditional relationships between customer and product attributes
• Log, Box-Cox and Yeo-Johnson transformations address skewed variables
• Binning continuous variables using quantiles, decision trees and business thresholds
• Feature selection using mutual information, L1 regularisation and permutation importance

Participants can build and compare baseline and regression models for continuous outcomes.

• Naive benchmarks establish performance floors before complex model development
• Linear regression assumptions tested through residual plots and diagnostics
• Ridge, Lasso and Elastic Net regularisation compared
• Decision tree regressors tuned for depth, leaves and split criteria
• Random forest regression handles nonlinear relationships and feature interactions
• MAE, RMSE, MAPE and R-squared interpreted against business requirements

Participants can develop binary and multiclass classification models with suitable decision thresholds.

• Logistic regression produces interpretable probability estimates for binary outcomes
• Decision trees classify records using transparent rule-based partitions
• Random forests reduce variance through bootstrap aggregation and feature sampling
• Gradient boosting models fitted with XGBoost-style iterative error correction
• Multiclass strategies compare one-versus-rest and multinomial classification approaches
• Classification thresholds selected using expected costs and operational capacity

Participants can develop calibrated classifiers for rare events and uneven class distributions.

• Class imbalance measured through prevalence, minority counts and sampling ratios
• Stratified splitting preserves class proportions across development datasets
• Class weights adjust learning objectives for asymmetric classification errors
• SMOTE and random undersampling compared with cost-sensitive learning
• Precision-recall curves assessed for fraud, default and incident detection
• Platt scaling and isotonic regression calibrate predicted probabilities

Participants can design valid testing procedures and diagnose model errors by segment.

• Holdout sets reserved for final unbiased model performance assessment
• K-fold cross-validation estimates stability across multiple training partitions
• Time-series splits prevent future observations contaminating historical training data
• Nested cross-validation separates parameter selection from performance estimation
• Confusion matrices analysed by segment, threshold and error consequence
• Residual and misclassification reviews identify systematic model failure patterns

Participants can tune model parameters efficiently while controlling overfitting and computational cost.

• Parameter spaces defined from algorithm behaviour and computational constraints
• Grid search evaluates structured combinations of selected hyperparameter values
• Random search samples broad parameter ranges more efficiently
• Bayesian optimisation prioritises promising configurations from prior trial results
• Early stopping limits boosting iterations when validation performance plateaus
• Validation results recorded with parameters, metrics, runtime and random seeds

Participants can select and evaluate ensemble methods for improved predictive performance.

• Bagging reduces model variance through repeated bootstrap training samples
• Random forests combine decorrelated trees for stable generalisation performance
• AdaBoost focuses successive learners on previously misclassified observations
• Gradient boosting optimises differentiable loss functions through additive trees
• Stacking combines diverse base learners using out-of-fold predictions
• Bias-variance trade-offs guide ensemble complexity and model selection

Participants can create and assess meaningful segments when labelled outcomes are unavailable.

• K-means clustering assigns observations to centroids using Euclidean distance
• Hierarchical clustering visualised through dendrograms and linkage methods
• DBSCAN identifies dense clusters while classifying isolated observations as noise
• Feature scaling and distance metrics selected before cluster development
• Silhouette scores, inertia and stability tests assess cluster quality
• Segment profiles translated into operational actions and treatment strategies

Participants can reduce high-dimensional data and identify unusual observations using suitable methods.

• Principal component analysis compresses correlated variables into orthogonal components
• Explained variance guides component retention and information-loss decisions
• t-SNE and UMAP visualise complex high-dimensional observation patterns
• Isolation Forest identifies anomalies through random partition path lengths
• Local Outlier Factor compares observations against neighbourhood density
• Anomaly investigation separates data errors, rare events and emerging risks

Participants can design and assess basic neural networks for structured prediction tasks.

• Feedforward neural networks map inputs through weighted hidden layers
• Activation functions: ReLU, sigmoid and tanh compared
• Backpropagation updates weights using gradients and learning-rate schedules
• Dropout, batch normalisation and weight decay reduce overfitting
• Embedding layers represent high-cardinality categorical variables efficiently
• Neural networks compared against tree ensembles for tabular datasets

Participants can explain model outputs, test fairness concerns and communicate limitations responsibly.

• Global feature importance summarises variables contributing to overall model predictions
• Partial dependence plots show average marginal feature effects
• Individual conditional expectation plots reveal heterogeneous prediction responses
• SHAP values attribute individual predictions to contributing feature values
• Fairness metrics compare selection rates, errors and calibration across groups
• Model limitations communicated through intended use, exclusions and uncertainty statements

Participants can document model evidence for technical review, approval and controlled use.

• Model inventory records ownership, purpose, materiality and approval status
• Development documentation captures data, features, algorithms and validation results
• Assumptions, limitations and known failure modes documented explicitly
• Independent validation plans test conceptual soundness and implementation accuracy
• Champion-challenger comparisons support model replacement and periodic review
• Model cards communicate intended use, performance and ethical considerations

Participants can design a controlled path from trained model to production service.

• Batch scoring and real-time inference matched to decision timing
• REST API contracts define inputs, outputs, authentication and error responses
• Feature parity checks prevent training-serving transformation inconsistencies
• Model versioning links deployed artefacts to approved development records
• Container deployment packages code, dependencies and runtime configuration
• Release gates require testing, approval, rollback and change records

Participants can define monitoring indicators, response thresholds and model review cycles.

• Data drift measured using population stability index and distribution tests
• Concept drift detected through declining predictive accuracy and calibration
• Production performance tracked against baseline metrics and service objectives
• Prediction distributions monitored for unexpected shifts and operational anomalies
• Alert thresholds assign escalation actions, accountable owners and response times
• Retraining triggers defined from drift, performance and policy changes

Participants can communicate model evidence and recommendations to technical, risk and business stakeholders.

• Executive model summaries present decision value, performance and key limitations
• Technical review packs include methodology, validation evidence and reproducibility records
• Threshold trade-offs presented using capacity, risk appetite and customer impact
• Model results translated into actions for operations, sales and risk teams
• Challenge questions anticipated from audit, compliance and independent validation
• Decision logs record approvals, conditions, exceptions and accountable owners

Participants can build, review and present a complete machine learning solution with validated performance, governance evidence and deployment recommendations.

• Business problem framed with target, decision owner and measurable success criteria
• Analytical dataset prepared through documented cleaning, transformation and feature engineering
• Candidate models compared using cross-validation, calibration and segment-level error analysis
• Selected model interpreted using SHAP values, fairness checks and limitations
• Deployment design specifies scoring pattern, controls, monitoring and retraining triggers
• Model review presentation defends recommendations before business, technical and risk stakeholders