Learning becomes durable when you make the system move.
Use working playgrounds for algorithms and evaluation, concept-specific challenges for retrieval practice, debugging exercises for diagnosis, and validated case studies for portfolio-style application.
Follow the data-science workflow from data to application.
Playgrounds follow a complete data-science workflow from data understanding through modelling, validation, robustness and applied experimentation.
Understand and prepare the evidence first.
Inspect the data, clean it safely, create useful features, reduce dimensionality, handle imbalance and learn to recognise leakage before model selection begins.
Dataset Playground
Built-in data, local CSV upload, schema profiling, missingness, duplicates, feature/target selection and reproducible splitting.
Open playground →Preprocessing Playground
Imputation, encoding, scaling and outlier handling fitted on training data and then applied to held-out rows.
Open playground →Feature Engineering Playground
Polynomial, interaction, binning, log, datetime and domain features with held-out performance comparison.
Open playground →Correlation Analysis & Feature Selection
Inspect Pearson/Spearman relationships, correlation matrices and stepwise correlation-based feature selection with redundancy control.
Open playground →Dimensionality Reduction Lab
Computed PCA plus explicitly conceptual t-SNE-style and UMAP-style neighbourhood projections in 2D/3D.
Open playground →Class Imbalance Lab
Prevalence, resampling, KNN, thresholds, confusion counts and PR-AUC with an untouched evaluation set.
Open playground →Data Leakage Simulator
Target, entity and future-information leakage compared directly with repaired validation workflows.
Open playground →Learn what the algorithms are doing.
Move from regression and clustering into decision boundaries, neural networks, optimisation, ensembles and step-by-step algorithm behaviour.
Regression Playground
Linear, polynomial, ridge and lasso regression with editable points, outliers, residuals and MAE/MSE/RMSE/R².
Open playground →Clustering Playground
Real K-Means, DBSCAN and hierarchical clustering with stepwise formation, linkage controls and cluster diagnostics.
Open playground →Decision Boundary Visualiser
Logistic regression, KNN, linear SVM, decision tree and random forest on editable 2D observations.
Open playground →Neural Network Playground
Configurable multilayer perceptron with backpropagation, loss curve, decision surface and probe activations.
Open playground →Gradient Descent Simulator
MSE loss surface with batch, stochastic and mini-batch updates, learning-rate experiments and optimisation path.
Open playground →Ensemble Learning Simulator
Bagging, AdaBoost, soft voting and stacking with base-vs-combined boundaries and metric comparison.
Open playground →Algorithm Step-by-Step Mode
Slow down KNN, K-Means, tree splitting, gradient descent, PCA and neural-network backpropagation.
Open playground →Build evidence for choosing a model.
Construct leakage-safe pipelines, compare model families fairly, tune without test-set peeking, inspect validation folds, diagnose bias/variance and retain experiment history.
ML Pipeline Builder
Dataset → clean → scale → engineer → select → model → validation → final test, with train-only fitted transformations.
Open playground →Model Comparison Lab
Train six model families on the same split, rank by validation evidence and test only the selected winner.
Open playground →Hyperparameter Playground
Model-specific validation curves with a deliberately locked final test until parameter selection is complete.
Open playground →Cross-Validation Playground
Holdout, K-Fold, Stratified K-Fold and Leave-One-Out with fold geometry, per-fold scores and variability.
Open playground →Overfitting / Underfitting Lab
Polynomial complexity, data size, noise, generalisation gaps and learning curves make bias–variance visible.
Open playground →Experiment History
Save experiment runs locally, filter and compare up to three experiments, then export the history as CSV or JSON.
Open playground →Ask whether the model is trustworthy.
Move beyond training score into confusion metrics, threshold behaviour, explanation, robustness, distribution shift and time-aware modelling.
Classification Metrics Simulator
Manipulate TP, FP, TN and FN directly and watch accuracy, precision, recall, specificity, F1 and balanced accuracy respond.
Open playground →Threshold Explorer
Move the decision threshold over fixed case scores and connect confusion counts with ROC and precision–recall operating points.
Open playground →Explainable AI Playground
Compare p-values, univariate AUC, permutation importance, RFE and local feature effects on the same model evidence.
Open playground →Noise & Robustness Playground
Inject label noise, feature noise, missingness or outliers into training/evaluation data and measure performance degradation.
Open playground →Train–Test Distribution Shift Simulator
Hold a trained separator fixed while covariate shift, concept drift or combined shift changes deployment evidence.
Open playground →Time-Series Playground
Control trend, seasonality and noise, preserve temporal order, forecast future observations and compare forecasting assumptions.
Open playground →Turn the components into an applied learning system.
Apply trained models to new cases, connect code with model behaviour, and solve goal-based ML challenges using the same evidence-first workflow.
Prediction Sandbox
Enter new feature values, obtain probabilities, inspect local feature effects and compare nearby training observations.
Open playground →Code ↔ Visualisation Synchronisation
Synchronise controls with a complete reusable Python script, step through its functions, and export the experiment as a .py file.
Open playground →Challenge Mode
Solve validation-scored ML goals, receive feedback, and unlock the final test only after the challenge criteria are met.
Open playground →Learn, experiment, or solve a challenge.
Every modern playground can use the shared mode switch. Learn keeps the educational framing visible, Experiment encourages free manipulation, and Challenge adds an applied goal with a direct route to scored challenges.
36 concept-specific checkpoints
Every flagship concept now has a direct Practice destination. No Lab-home fallback.
Open challenge library →12 validated case studies
Download data, calculate answers, justify decisions, validate tasks and export a local submission report.
Open projects →