Data Analysis and Machine Learning Using R Training Course

Participants can import structured data into R and diagnose material quality issues before analysis.

• CSV, Excel and database extracts imported with explicit data types
• Inspecting dimensions, variable classes, unique values and duplicate records
• Missingness profiles separated by variable, record and source system
• Range checks identify impossible dates, amounts, codes and categories
• Reconciling record counts against source extracts and control totals
• Quality findings documented in an auditable data assessment log

Participants can transform and combine raw data into analysis-ready tables using tidyverse functions.

• filter, select and mutate operations create scoped analytical datasets
• Grouped summaries calculate counts, rates, totals and distribution statistics
• Joining tables with primary keys, foreign keys and match checks
• pivot_longer and pivot_wider reshape repeated measurement data correctly
• String and date transformations standardise categories and time fields
• Pipeline scripts separate raw, intermediate and curated data outputs

Participants can investigate distributions, relationships and anomalies using targeted statistical graphics.

• Histograms, density plots and boxplots assess continuous variable distributions
• Bar charts and proportion tables examine categorical variable composition
• Scatterplots reveal nonlinear relationships, outliers and heteroscedasticity patterns
• Faceting compares distributions across regions, products or customer segments
• Correlation matrices screen associations among numeric predictor variables
• ggplot2 scales, labels and themes produce readable analytical graphics

Participants can apply confidence intervals and hypothesis tests while stating their assumptions and limitations.

• Sampling distributions explain uncertainty in estimated means and proportions
• Confidence intervals communicate plausible ranges around sample estimates
• Two-sample tests compare groups using appropriate assumptions and alternatives
• Chi-square tests assess association between categorical variables and outcomes
• Effect sizes distinguish practical significance from small p-values
• Multiple testing risks considered when screening many variables simultaneously

Participants can fit, diagnose and interpret linear regression models for continuous targets.

• Ordinary least squares estimates relationships between predictors and outcomes
• Coefficient interpretation distinguishes units, reference categories and conditional effects
• Residual plots assess linearity, constant variance and influential observations
• Variance inflation factors identify problematic multicollinearity among predictors
• Interaction terms represent conditional relationships between explanatory variables
• Prediction intervals distinguish individual forecasts from mean response estimates

Participants can build and interpret logistic regression models for binary decision outcomes.

• Binary target encoding defines positive classes and event direction
• Logistic regression models log-odds for observed binary outcomes
• Odds ratios communicate predictor effects with suitable reference categories
• Predicted probabilities support threshold-based operational decision rules
• Confusion matrices separate true positives, false positives and misses
• ROC curves and AUC compare discrimination across classification models

Participants can create, transform and select predictors without contaminating model evaluation.

• Derived ratios, lags and aggregates encode operational behaviour patterns
• Log transformations reduce skewness in monetary and count variables
• Dummy variables represent nominal categories for regression-based models
• Scaling and centring prepare numeric predictors for distance-based methods
• Missing-value imputation fitted within training data prevents information leakage
• Feature recipes document repeatable preprocessing steps for new records

Participants can estimate out-of-sample model performance using appropriate resampling designs.

• Training, validation and test sets preserve an unbiased final assessment
• K-fold cross-validation estimates performance across repeated data partitions
• Stratified sampling preserves event prevalence in classification partitions
• Time-based splits respect temporal ordering in forecasting and risk data
• Hyperparameter tuning compares candidate settings without using test data
• Selection criteria balance predictive performance, stability and operational usability

Participants can train and tune tree-based models for nonlinear prediction problems.

• Recursive partitioning creates interpretable decision rules from predictor splits
• Tree depth and minimum node size control model complexity
• Random forests reduce variance through bootstrap aggregation and feature sampling
• Gradient boosting improves predictions through sequential residual fitting
• Variable importance ranks predictors while recognising correlated-feature limitations
• Partial dependence plots show average predictor effects on predictions

Participants can assess distance-based and margin-based classifiers against baseline models.

• k-nearest neighbours classify records using local feature-space proximity
• Distance metrics require scaled predictors and meaningful feature selection
• Support vector machines maximise class separation using decision margins
• Kernel functions represent nonlinear boundaries without explicit transformations
• Tuning neighbour counts, cost and kernels uses cross-validation results
• Benchmark comparisons identify when simpler models remain preferable

Participants can design classification evaluation and threshold rules for rare-event outcomes.

• Prevalence establishes the baseline difficulty of rare-event prediction
• Precision, recall and F1 score assess minority-class model performance
• Precision-recall curves provide clearer evidence than ROC under imbalance
• Class weights alter training loss for unequal error consequences
• Oversampling and undersampling applied only within resampling partitions
• Decision thresholds reflect investigation capacity and false-positive costs

Participants can create and assess useful clusters when no labelled outcome is available.

• Standardised features prepare mixed-scale data for clustering algorithms
• K-means clustering partitions observations around within-cluster similarity
• Hierarchical clustering displays nested groups through dendrogram structures
• Silhouette scores assess cohesion and separation among candidate clusters
• Cluster profiles compare behaviours, needs and value measures
• Segmentation labels translated into actionable service or treatment strategies

Participants can build baseline forecasts and validate them using time-respecting evaluation methods.

• Trend, seasonality and calendar effects identified through time plots
• Lagged predictors capture persistence in operational and financial series
• Moving averages provide transparent baseline forecasting methods
• ARIMA models represent autoregressive, differencing and moving-average components
• Rolling-origin resampling evaluates forecasts across successive time periods
• MAE, RMSE and MAPE compared against business forecasting requirements

Participants can explain model behaviour, identify material risks and document responsible-use controls.

• Coefficient tables and marginal effects explain regression model behaviour
• Permutation importance tests sensitivity of predictions to each feature
• Individual conditional expectation plots show case-level prediction variation
• Proxy variables assessed for potential protected-group discrimination risks
• Subgroup performance checks reveal unequal error rates across populations
• Model cards record purpose, data, metrics, limits and approvals

Participants can produce an auditable analytical report combining R code, outputs and recommendations.

• Quarto documents combine narrative, code, tables and graphics
• Chunk options control code visibility, warnings and figure presentation
• Parameters support repeatable reports for different periods or regions
• Versioned inputs and package dependencies support reproducible reruns
• Executive summaries state findings, uncertainty and decision implications
• Technical appendices preserve methods, assumptions and validation evidence

Participants can prepare a model handover plan covering scoring, monitoring and ownership.

• Scoring specifications define required inputs, transformations and prediction outputs
• Data refresh schedules identify timing, ownership and quality controls
• Performance monitoring tracks drift, calibration and outcome availability
• Champion-challenger comparisons support controlled model replacement decisions
• Approval records document accountable owners, review dates and restrictions
• Implementation plans distinguish analytical prototypes from production services

Participants can build, review and present a complete R-based analysis with validated model recommendations.

• Decision brief defines stakeholders, target outcome and measurable success criteria
• Data pipeline imports, validates and transforms source records reproducibly
• Exploratory findings identify patterns, anomalies and candidate predictor variables
• Competing models compared through appropriate resampling and performance metrics
• Interpretation addresses drivers, uncertainty, fairness and operational constraints
• Quarto report presents recommendations, monitoring requirements and implementation actions