Cortex Model Atlas

Models explained like engineering tools not buzzwords.

Understand what each model does, when it is useful, where it fails, what to measure and how it behaves in production.

1Linear models
1Tree ensembles
2Boosted trees
1Clustering
Linear models

Logistic Regression

A probabilistic classification baseline that is transparent, fast and surprisingly competitive.

PR-AUCROC-AUCLog loss
Binary classificationStudy model →
Tree ensembles

Random Forest

An ensemble of decorrelated decision trees that captures nonlinear interactions with limited preprocessing.

PR-AUCROC-AUCRMSE
Tabular classificationStudy model →
Boosted trees

Gradient Boosting

Sequential trees correct residual errors and often dominate structured-data benchmarks.

PR-AUCLog lossRMSE
High-quality tabular predictionStudy model →
Clustering

K-Means

Partitions observations around centroids and provides a simple baseline for segmentation.

SilhouetteInertiaStability
Exploratory segmentationStudy model →
Dimensionality reduction

PCA

Rotates data into orthogonal directions that capture maximum variance.

Explained varianceReconstruction error
CompressionStudy model →
Neural networks

Multilayer Perceptron

A stack of learned affine transformations and nonlinearities that approximates complex functions.

LossAccuracy/F1Calibration
General nonlinear mappingStudy model →
Neural networks

Convolutional Neural Network

Learns local spatial filters and hierarchical visual features.

Accuracy/F1mAPLatency
ImagesStudy model →
Sequence models

Transformer

Uses attention to model relationships between tokens without recurrent state.

PerplexityTask accuracyLatency
LanguageStudy model →
Information retrieval

BM25

A strong lexical ranking algorithm balancing term frequency, rarity and document length.

MRRnDCG@kPrecision@k
SearchStudy model →
Information retrieval

TF-IDF

Represents text by upweighting terms frequent in a document but rare across the corpus.

Cosine similarityRetrieval precisionClassifier F1
Text similarityStudy model →
Reinforcement learning

Q-Learning

An off-policy temporal-difference method that learns action values from reward transitions.

Average rewardRegretPolicy stability
Discrete decision policiesStudy model →
Reinforcement learning

Contextual Bandit

Chooses among actions using context while learning from immediate reward.

RegretRewardAction coverage
Content selectionStudy model →
Bayesian bandits

Thompson Sampling

Samples action quality from posterior beliefs to balance exploration and exploitation.

RegretPosterior calibrationReward
A/B allocationStudy model →
Decision theory

Markov Decision Process

A formal model of states, actions, transition probabilities, rewards and discounting.

ReturnPolicy valueState coverage
Sequential decision problemsStudy model →
Probabilistic models

Naive Bayes

Uses Bayes rule with conditional independence assumptions for fast classification.

F1Log lossCalibration
Text classificationStudy model →
Margin methods

Support Vector Machine

Finds a maximum-margin decision boundary and can use kernels for nonlinear separation.

F1Margin violationsLatency
Medium datasetsStudy model →
Anomaly detection

Isolation Forest

Detects anomalies by how quickly random trees isolate a point.

Precision@kRecall@kAlert rate
Outlier detectionStudy model →
Time series

ARIMA

Models autoregressive structure, differencing and moving-average residuals.

MAERMSEMAPE
Classical forecastingStudy model →
Boosted trees

XGBoost

Regularized gradient boosting engineered for strong tabular performance and efficient training.

PR-AUCNDCGRMSE
Tabular ranking/classification/regressionStudy model →
Graph learning

Graph Neural Network

Learns node/edge representations by passing messages across graph neighborhoods.

Hits@kMRRNode F1
Knowledge graphsStudy model →