> ML_ALGORITHMS_ATLAS_v1.0
Algorithms & Method Families
100 qualified algorithmic method families across 9 disciplines: Classical Supervised, Unsupervised, Time Series, Anomaly Detection, Deep Learning, Recommenders, Causal Inference, Reinforcement Learning, and Specialized Methods.
Ordinary Least Squares Linear Regression
Directly readable regression coefficients represent marginal impact per unit change.
Logistic Regression (Logit)
Exponentiated coefficients directly reflect Odds Ratios per unit predictor increase.
Regularized Linear Models (Ridge, Lasso, ElasticNet)
Lasso enforces exact sparsity, setting non-informative feature weights to zero.
CART Decision Trees
Can be visually inspected as human-executable IF-THEN logic trees.
Random Forests
Aggregated ensemble cannot be read as a single tree; relies on MDI, MDA, and SHAP values.
Extremely Randomized Trees (Extra Trees)
Ensemble evaluated via feature importances and TreeSHAP.
Gradient Boosted Decision Trees (GBDT)
TreeSHAP provides mathematically rigorous local and global feature attributions.
Modern Histogram & Symmetric Boosting (XGBoost / LightGBM / CatBoost)
Fast TreeSHAP algorithms compute exact Shapley values in milliseconds.
Support Vector Machines (SVM)
Non-linear kernels (RBF, Polynomial) operate in infinite-dimensional space; weights cannot be directly inspected.
k-Nearest Neighbors (k-NN)
Predictions can be explained by presenting the exact nearest historical neighbor instances.
Naive Bayes Classifiers (Gaussian, Multinomial, Bernoulli)
Individual log-likelihood contributions per word or feature can be summed and inspected.
Discriminant Analysis (LDA / QDA)
Linear discriminant projections provide clear orthogonal class separation axes.
K-Means & K-Means++ Clustering
Centroids directly represent the average feature vector of all members in each cluster.
Principal Component Analysis (PCA & TruncatedSVD)
Principal axes are linear combinations of features whose loadings indicate exact contribution.
t-Distributed Stochastic Neighbor Embedding (t-SNE)
Provides high-contrast 2D/3D visualizations of clusters, though inter-cluster distances are not calibrated.
Uniform Manifold Approximation and Projection (UMAP)
Preserves significantly more global topological structure than t-SNE while maintaining tight local clusters.
HDBSCAN & DBSCAN Density-Based Clustering
Explicitly outputs an outlier/noise class (-1) without forcing non-conforming points into spurious clusters.
Hierarchical Agglomerative Clustering (Ward / Complete Linkage)
Dendrogram tree visualization reveals the full multi-scale hierarchy of data groupings.