Data Science and Machine Learning Training Course

Participants can organise reproducible Python notebooks and project files for analytical work.

• Python environments managed with virtual environments, requirements files and package versions
• Notebook structure separating imports, parameters, transformations, outputs and commentary
• Functions created for repeatable data cleaning and feature generation tasks
• Git commits used to record analytical changes and modelling decisions
• Random seeds controlled for repeatable sampling, splitting and model training
• Project directories organised around raw data, processed data, code and reports

Participants can extract, join and validate source data for a modelling dataset.

• SQL joins selected according to entity grain and relationship cardinality
• Common table expressions used to stage readable multi-step data transformations
• Window functions calculate customer recency, transaction counts and rolling measures
• Entity keys checked for duplicates, missing identifiers and unexpected relationship changes
• Point-in-time extracts prevent future information entering historical training records
• Data dictionaries capture field meaning, source ownership, refresh frequency and sensitivity

Participants can profile data quality and identify patterns relevant to model design.

• Missingness patterns examined using counts, proportions and indicator variables
• Distribution plots reveal skewness, outliers, truncation and implausible values
• Summary statistics compared across time periods, segments and source systems
• Correlation analysis distinguishes linear association from multicollinearity concerns
• Data leakage identified through timing, proxy variables and post-outcome fields
• Exploratory findings documented as testable hypotheses for subsequent modelling

Participants can build repeatable data cleaning pipelines suitable for model training.

• pandas methods filter records, convert types and standardise categorical values
• Duplicate records resolved using business keys, timestamps and survivorship rules
• Missing numeric values imputed with median, model-based or domain-informed methods
• Categorical missingness treated as separate information where operationally meaningful
• Outlier handling uses winsorisation, capping, transformation or investigation decisions
• scikit-learn pipelines combine preprocessing steps with reproducible model fitting

Participants can apply essential statistical concepts when interpreting model results.

• Sampling distributions explain variation between training samples and model estimates
• Confidence intervals communicate uncertainty around averages, rates and performance measures
• Bayes' theorem relates prior probability, evidence and posterior probability
• Bias-variance trade-off explains underfitting, overfitting and generalisation error
• Regularisation penalises unnecessary model complexity through L1 and L2 terms
• Statistical significance distinguished from business relevance and decision usefulness

Participants can create and select meaningful features from structured operational data.

• Ratio features combine related measures while preserving meaningful denominator controls
• Date features derive seasonality, tenure, recency and elapsed-time measures
• Categorical encoding uses one-hot, ordinal, frequency and target encoding approaches
• Interaction terms represent conditional effects between customer, product and channel attributes
• Feature scaling applies standardisation, normalisation and log transformation appropriately
• Feature selection uses domain logic, mutual information and regularised model coefficients

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

• Linear regression models continuous outcomes using estimated predictor coefficients
• Residual diagnostics assess linearity, heteroscedasticity, normality and influential observations
• Ridge and lasso regression control coefficient size and feature selection
• Mean absolute error compared with root mean squared error
• Prediction intervals distinguish expected estimates from plausible outcome ranges
• Regression outputs translated into forecast assumptions and operational planning actions

Participants can build classification models and select decision thresholds for binary outcomes.

• Logistic regression estimates event probabilities from predictor variables and coefficients
• Confusion matrices classify true positives, false positives and related errors
• ROC curves compare discrimination performance across alternative decision thresholds
• Precision-recall curves assess rare-event models where positive cases are scarce
• Probability calibration checks whether predicted risks match observed event rates
• Threshold selection reflects intervention capacity, error costs and policy constraints

Participants can compare tree-based models and tune ensembles for predictive performance.

• Decision trees split observations using impurity reduction and stopping criteria
• Tree depth, minimum samples and pruning control overfitting behaviour
• Random forests aggregate decorrelated trees through bootstrap sampling and feature selection
• Gradient boosting improves weak learners through sequential residual correction
• XGBoost parameters tune learning rate, depth, subsampling and regularisation
• Feature importance interpreted cautiously alongside correlation and leakage checks

Participants can select models using defensible validation designs and comparative evidence.

• Train-validation-test splits preserve an untouched final performance assessment dataset
• Stratified sampling maintains outcome proportions for imbalanced classification datasets
• K-fold cross-validation estimates performance variation across multiple training partitions
• Time-series splits preserve chronological order for forecasting and event prediction
• Grid search and random search tune hyperparameters against validation metrics
• Model selection balances accuracy, interpretability, latency and maintenance requirements

Participants can address rare-event classification using appropriate sampling and cost methods.

• Class prevalence assessed before selecting metrics, sampling methods or thresholds
• Class weights increase the learning penalty for minority-class prediction errors
• Random undersampling reduces majority examples while protecting representative variation
• SMOTE generates minority examples between nearby observations in feature space
• Precision at k measures investigation yield within limited operational capacity
• Cost matrices translate classification errors into financial and service consequences

Participants can create and assess useful segments where labelled outcomes are unavailable.

• K-means clustering groups observations around iteratively updated centroids
• Feature scaling prevents large-unit variables dominating distance-based clustering calculations
• Elbow and silhouette methods assess plausible cluster number choices
• Hierarchical clustering reveals nested segment structures using linkage distances
• Principal component analysis reduces correlated dimensions for visual inspection
• Segment profiles connect statistical clusters to actionable customer or operational groups

Participants can prepare time-series data and evaluate forecasting approaches for operational planning.

• Time-series decomposition separates trend, seasonality and residual variation components
• Lag features represent prior observations for supervised forecasting models
• Rolling windows calculate moving averages, volatility and recent demand measures
• ARIMA models specify autoregressive, differencing and moving-average components
• Forecast accuracy uses MAE, RMSE, MAPE and weighted error measures
• Backtesting compares forecasts against historical periods using chronological validation

Participants can prepare text data and build baseline text classification models.

• Text normalisation applies lowercasing, tokenisation, punctuation removal and lemmatisation
• Stop-word decisions reflect domain language, classification purpose and interpretability requirements
• Bag-of-words vectors represent document terms through count-based feature matrices
• TF-IDF weighting increases distinctive terms while reducing common word influence
• N-grams capture short phrases relevant to complaints, requests and intent detection
• Baseline classifiers assess text categorisation before more complex language models

Participants can explain model decisions and assess fairness risks before deployment.

• Global feature importance identifies variables associated with overall model predictions
• Partial dependence plots show average predicted response across feature values
• SHAP values explain individual predictions through additive feature contributions
• Protected characteristics distinguished from proxy variables and legitimate risk factors
• Fairness metrics compare selection rates, error rates and calibration between groups
• Model cards document intended use, limitations, evaluation results and governance owners

Participants can specify a practical deployment design for a trained machine learning model.

• Batch scoring schedules predictions for periodic campaigns, prioritisation and planning
• Real-time APIs support low-latency decisions within transactional application workflows
• Feature stores align training features with production scoring inputs
• Model artefacts version code, parameters, data references and trained outputs
• Acceptance tests check schema, prediction ranges and business-rule compatibility
• Human review routes uncertain predictions and exceptions to accountable decision-makers

Participants can define monitoring controls, retraining triggers and model governance responsibilities.

• Data drift measures changes in production feature distributions over time
• Concept drift identifies declining relationships between predictors and target outcomes
• Performance monitoring tracks accuracy, calibration, latency and operational intervention rates
• Population stability index flags material changes in scored population characteristics
• Retraining triggers define evidence thresholds, approvals and rollback arrangements
• Model governance assigns ownership, review frequency, documentation and audit records

Participants can present model findings, limitations and recommendations to decision-makers.

• Executive summaries state the decision, evidence, recommendation and residual uncertainty
• Performance charts compare baseline, candidate model and business-rule alternatives
• Error analysis identifies affected groups, failure patterns and operational consequences
• Scenario analysis estimates expected value under alternative thresholds and capacity limits
• Technical documentation records data lineage, assumptions, parameters and validation evidence
• Recommendations distinguish immediate actions, controlled pilots and further investigation needs

Participants can build, review and present a complete machine learning project from problem definition through deployment controls.

• Problem charter defines decision users, target outcome, constraints and success measures
• Prepared dataset combines documented extraction, cleaning, feature engineering and quality checks
• Candidate models compared through cross-validation, metrics, calibration and error analysis
• Selected model justified through performance, interpretability, cost and implementation requirements
• Model card records fairness evidence, limitations, intended use and approval conditions
• Deployment plan specifies scoring method, monitoring metrics, retraining triggers and owners