Machine Learning with Python Training Course

Participants can organise a Python machine learning project for repeatable analysis and review.

• Jupyter notebooks structured with assumptions, methods, outputs and decisions
• Virtual environments isolate package versions and project dependencies
• Python modules separate reusable functions from exploratory notebook code
• Random seeds control reproducibility in sampling and model fitting
• File paths managed with pathlib instead of machine-specific strings
• Git commits record meaningful changes to data and modelling code

Participants can load, inspect and profile tabular data before modelling decisions are made.

• CSV, Excel and Parquet files loaded with appropriate data types
• DataFrame dimensions, columns and sample records inspected systematically
• Missingness patterns profiled by variable, record and business segment
• Duplicate records identified using keys, timestamps and field comparisons
• Date fields parsed into timezone-aware datetime representations
• Data dictionaries link source fields to business definitions and owners

Participants can clean source data while preserving an auditable record of transformations.

• Validity checks identify impossible values, ranges and category codes
• Missing numeric values treated using median, model-based or indicator methods
• Categorical inconsistencies standardised through mapping tables and controlled vocabularies
• Outliers assessed using domain thresholds, IQR rules and distribution plots
• Cleaning functions retain raw fields alongside transformed analytical variables
• Transformation logs document assumptions, exceptions and unresolved data issues

Participants can investigate distributions, relationships and data risks that influence model design.

• Distribution plots reveal skewness, multimodality, truncation and unusual observations
• Grouped summaries compare outcomes across products, regions and customer segments
• Correlation matrices identify linear associations and redundant numeric predictors
• Cross-tabulations expose category frequencies, sparsity and outcome concentration
• Time-series plots detect drift, seasonality and data collection changes
• Target distributions checked before selecting classification or regression approaches

Participants can apply core statistical concepts when interpreting model estimates and validation results.

• Sampling distributions explain uncertainty in observed model performance estimates
• Bias and variance describe underfitting, overfitting and generalisation error
• Conditional probability supports interpretation of classification outcomes and base rates
• Covariance and correlation distinguish association from predictive usefulness
• Confidence intervals communicate uncertainty around performance and model coefficients
• Statistical significance distinguished from practical decision-making relevance

Participants can create, transform and select model features without contaminating validation data.

• Date features extract recency, frequency, duration and calendar effects
• Ratios and interaction terms represent business relationships between variables
• One-hot encoding converts nominal categories into model-ready indicator variables
• Ordinal encoding preserves meaningful rank order within categorical fields
• Log and Box-Cox transformations address highly skewed positive variables
• Feature selection uses domain reasoning, mutual information and regularisation

Participants can create defensible data partitions and validation schemes for machine learning models.

• Holdout test sets reserved until final model selection decisions
• Stratified sampling preserves class proportions across classification partitions
• K-fold cross-validation estimates performance across repeated training subsets
• Time-based splits prevent future information entering historical predictions
• Group-based splits prevent related records appearing in both partitions
• Data leakage checks cover targets, timestamps, aggregates and preprocessing steps

Participants can build and assess regression models for continuous business outcomes.

• Ordinary least squares regression establishes an interpretable baseline model
• Residual plots assess heteroscedasticity, nonlinearity and influential observations
• Mean absolute error and RMSE compare continuous prediction performance
• R-squared interpreted alongside error metrics and operational tolerances
• Ridge and Lasso regression control coefficient instability and feature selection
• Prediction intervals distinguish expected values from plausible outcome ranges

Participants can train and assess binary and multiclass classification models for operational decisions.

• Logistic regression estimates class probabilities from predictor variables
• Confusion matrices separate true positives, false positives and missed cases
• Precision, recall and F1 score reflect different error priorities
• ROC curves compare discrimination across possible decision thresholds
• Precision-recall curves assess rare-event classification performance more clearly
• Threshold selection incorporates workload capacity, risk appetite and intervention costs

Participants can build preprocessing and modelling pipelines that prevent inconsistent transformations.

• ColumnTransformer applies separate transformations to numeric and categorical features
• SimpleImputer fits missing-value rules using training data only
• StandardScaler prepares scale-sensitive algorithms for reliable model fitting
• Pipeline objects combine preprocessing, estimators and cross-validation procedures
• Feature names traced through transformations for model review and interpretation
• Joblib serialisation saves fitted pipelines with versioned model artefacts

Participants can use tree-based models to capture nonlinear patterns and feature interactions.

• Decision tree splits selected using Gini impurity or variance reduction
• Maximum depth and minimum samples control tree complexity
• Random forests aggregate decorrelated trees to reduce prediction variance
• Out-of-bag estimates provide internal validation for random forest models
• Permutation importance tests performance loss after feature values are shuffled
• Partial dependence plots show average feature effects on predictions

Participants can tune boosted models systematically while controlling overfitting and computational cost.

• Gradient boosting fits successive trees to reduce residual prediction error
• Learning rate and estimator count jointly control boosting behaviour
• GridSearchCV compares defined parameter combinations using cross-validation
• RandomizedSearchCV samples wider hyperparameter spaces within fixed computing budgets
• Validation scores compared with training scores to detect overfitting
• Search results recorded with parameters, metrics, runtime and selected rationale

Participants can address rare events and identify unusual observations using suitable modelling techniques.

• Class prevalence quantified before selecting resampling or weighting strategies
• Class weights increase attention to minority outcomes during model fitting
• Random undersampling and oversampling evaluated within cross-validation folds
• SMOTE generates synthetic minority examples from nearest-neighbour relationships
• Isolation Forest identifies unusual records without labelled anomaly examples
• Rare-event results reviewed using precision, recall and alert volumes

Participants can build and assess clusters for segmentation and exploratory pattern discovery.

• K-means clustering groups records using distances from learned centroids
• Feature scaling prevents large-unit variables dominating cluster assignments
• Elbow plots compare within-cluster variation across candidate cluster counts
• Silhouette scores assess separation and cohesion among identified clusters
• Hierarchical clustering produces dendrograms for alternative segmentation structures
• Cluster profiles translate statistical groups into operationally meaningful segments

Participants can reduce high-dimensional data while retaining useful variation for analysis and modelling.

• Principal component analysis transforms correlated variables into orthogonal components
• Explained variance guides selection of retained principal components
• Standardisation precedes PCA when features use different measurement scales
• Component loadings identify original variables contributing to each component
• PCA supports visual exploration of high-dimensional customer and transaction data
• Dimensionality reduction assessed against interpretability and predictive performance

Participants can explain model behaviour and identify governance issues before operational use.

• Feature importance ranked with permutation methods and tree-based measures
• SHAP values explain individual predictions and aggregate feature contributions
• Partial dependence plots test plausible directional relationships in model behaviour
• Fairness checks compare error rates across relevant protected groups
• Model cards record intended use, limits, data sources and performance
• Human review points defined for high-impact or low-confidence decisions

Participants can prepare a practical plan for releasing, monitoring and reviewing a machine learning model.

• Batch scoring and real-time scoring matched to decision latency requirements
• Input validation checks schemas, ranges, categories and missing fields
• Data drift monitored through population stability index and distribution tests
• Concept drift identified when outcome relationships change over time
• Performance monitoring compares recent predictions against realised outcomes
• Retraining triggers specify ownership, evidence thresholds and approval steps

Participants can build, review and present a complete machine learning solution with documented recommendations.

• Business question translated into target, features, constraints and success criteria
• Source data profiled, cleaned and transformed using reproducible Python code
• Candidate models compared through cross-validation and decision-relevant metrics
• Selected pipeline tested against held-out data and plausible error scenarios
• Interpretation findings documented with limitations, fairness checks and governance controls
• Stakeholder presentation communicates recommendation, implementation plan and monitoring measures