| No. | Algorithm | Character | Unit | Topic |
|---|---|---|---|---|
| 1.1 | Naive Bayes Classifier | Probabilistic • Generative | Supervised Learning | 1.1 Classification |
| 1.2 | Logistic Regression | Linear • Discriminative | Supervised Learning | 1.1 Classification |
| 1.3 | K-Nearest Neighbors (KNN) | Instance-Based • Non-Parametric | Supervised Learning | 1.1 Classification |
| 1.4 | Support Vector Machine (SVM) | Margin Maximisation • Kernel Methods | Supervised Learning | 1.1 Classification |
| 1.5 | Decision Tree (Classification) | Tree-Based • Interpretable | Supervised Learning | 1.1 Classification |
| 1.6 | Linear Regression | Parametric • Closed-Form | Supervised Learning | 1.2 Regression |
| 1.7 | Polynomial Regression | Non-Linear • Feature Engineering | Supervised Learning | 1.2 Regression |
| 1.8 | Ridge Regression (L2 Regularisation) | Regularisation • Shrinkage | Supervised Learning | 1.2 Regression |
| 1.9 | Lasso Regression (L1 Regularisation) | Regularisation • Feature Selection | Supervised Learning | 1.2 Regression |
| 1.10 | Random Forest Regression | Ensemble • Bagging | Supervised Learning | 1.2 Regression |
| 2.1 | K-Means Clustering | Centroid-Based • Partitional | Unsupervised Learning | 2.1 Clustering |
| 2.2 | DBSCAN | Density-Based • Arbitrary Shapes | Unsupervised Learning | 2.1 Clustering |
| 2.3 | Hierarchical / Agglomerative Clustering | Hierarchical • Dendrogram | Unsupervised Learning | 2.1 Clustering |
| 2.4 | Apriori Algorithm | Frequent Itemsets • Association Rules | Unsupervised Learning | 2.2 Association Rule Learning |
| 2.5 | FP-Growth Algorithm | Tree-Based Mining • Scalable | Unsupervised Learning | 2.2 Association Rule Learning |
| 2.6 | Isolation Forest | Anomaly Detection • Tree Ensemble | Unsupervised Learning | 2.3 Anomaly Detection |
| 3.1 | Self-Training | Pseudo-Labelling • Iterative | Semi-Supervised Learning | 3.1 Inductive Methods |
| 3.2 | Co-Training | Multi-View • Ensemble | Semi-Supervised Learning | 3.1 Inductive Methods |
| 3.3 | Label Propagation | Graph-Based • Manifold | Semi-Supervised Learning | 3.2 Transductive Methods |
| 4.1 | Q-Learning | Off-Policy • Temporal Difference | Reinforcement Learning | 4.1 Model-Free Methods |
| 4.2 | Policy Gradient (REINFORCE) | On-Policy • Continuous Actions | Reinforcement Learning | 4.1 Model-Free Methods |
| 4.3 | Dyna-Q | Model-Based • Planning | Reinforcement Learning | 4.2 Model-Based Methods |
| 4.4 | Value Iteration (Dynamic Programming) | Model-Based • Exact | Reinforcement Learning | 4.2 Model-Based Methods |
| Unit | Additionally assumes |
|---|---|
| 1 — Supervised | Least squares; the idea of a likelihood; matrix inversion and rank. |
| 2 — Unsupervised | Distance metrics; set notation; conditional probability for lift. |
| 3 — Semi-Supervised | Unit 1 classifiers; graphs and adjacency matrices; eigen-intuition helps. |
| 4 — Reinforcement | Markov chains; expectation over trajectories; geometric series for discounting. |
Stated plainly, because a set of notes that claims to be complete should say where it stops. Each of these is a real gap rather than a deliberate exclusion:
| Missing | Why it matters |
|---|---|
| Dimensionality reduction — PCA, SVD, t-SNE, UMAP | Unit 2 covers clustering, association rules and anomaly detection, but no latent-variable methods. Ridge regression already leans on the SVD, and Label Propagation on the manifold assumption these would explain. |
| Gaussian Mixture Models and EM | The probabilistic generalisation of K-Means. It is what makes K-Means' "spherical, equally-sized clusters" limitation understandable rather than merely stated. |
| Boosting — AdaBoost, Gradient Boosting, XGBoost | Random Forest covers bagging (variance reduction). Without boosting (bias reduction), the ensemble idea is only half present. |
| Evaluation and model selection | Cross-validation, the bias–variance decomposition, data leakage, class imbalance, calibration and metric choice are used throughout these notes but never taught in their own right. |
| Preprocessing and feature engineering | "Scale your features first" is asserted in three separate algorithms and explained in none. |
| SARSA | Q-learning is labelled off-policy, but with no on-policy method beside it the label has nothing to contrast against. |
| Neural networks | DQN, CNN embeddings and function approximation are all referred to as though they had been introduced. They have not. |
$ in the text is always currency, never a delimiter.